-
A/B testing, also called split testing or controlled experimentation, is the practice of randomly assigning users (or sessions, or visits) to two or more variants of an experience and measuring the...
-
Adam, short for Adaptive Moment Estimation, is the adaptive optimization algorithm introduced by Kingma and Ba in 2014 that has become the default optimizer for deep learning including LLM pretrain...
-
AdamW is a variant of the Adam optimizer introduced by Loshchilov and Hutter in 2017 that decouples weight decay from the gradient-based parameter updates, producing better generalization and train...
-
Adapter layers are a PEFT technique introduced by Houlsby et al. in 2019 that inserts small bottleneck feed-forward modules between the frozen layers of a pretrained transformer , training only the...
-
Adversarial examples are inputs deliberately crafted to cause an AI model to produce incorrect or harmful outputs, often by introducing perturbations imperceptible to humans but disruptive to the m...
-
Agent handoff is the pattern where one LLM agent delegates a conversation or task to another agent better suited to handle it — analogous to a customer-service representative transferring a call to...
-
Agent memory is the persistent state that an agentic LLM system maintains across interactions, enabling continuity, learning, and personalization beyond a single conversation. Common memory types i...
-
An agent supervisor is the meta-agent in a multi-agent system whose role is not to do the work directly but to plan, dispatch, monitor, and synthesize across subordinate agents — essentially playin...
-
The Blueprint for an AI Bill of Rights is a non-binding policy framework published by the White House Office of Science and Technology Policy (OSTP) in October 2022, articulating five principles fo...
-
An AI Impact Assessment, also called an Algorithmic Impact Assessment or AIA, is a structured documentation and review process that evaluates the potential impacts of an AI system on individuals, g...
-
An AI taxonomy is the structured hierarchical classification system applied to AI-related content — documents, model outputs, retrieved chunks, conversations, user queries — that enables filtering,...
-
AI watermarking is the family of techniques for embedding detectable signatures in AI-generated content — text, images, audio, video — that allow downstream verifiers to determine that the content ...
-
Algorithmic accountability is the policy and engineering principle that organizations deploying automated decision-making systems must be answerable for those decisions — including the ability to e...
-
ALiBi, short for Attention with Linear Biases, is a positional encoding technique introduced by Press, Smith, and Lewis in a 2021 paper that adds linear-distance penalties directly to attention sco...
-
Alignment tax is the term used to describe the trade-off between safety properties and raw capability that often emerges when LLMs undergo RLHF , refusal training , Constitutional AI , or other saf...
-
AlpacaEval is an automated evaluation framework for chat-tuned LLMs released by Stanford's Tatsu Lab in 2023, scoring models by win rate against a reference model (originally text-davinci-003) on a...
-
Alt text, the alt attribute on HTML image elements, is the textual alternative that conveys an image's information to users who cannot see it — primarily screen-reader users who hear the alt text s...
-
Approval Chain Governance is an enterprise discipline within Workflow & Approvals focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and monit...
-
ARC, short for the AI2 Reasoning Challenge, is a benchmark introduced by Allen Institute for AI in 2018 containing 7,787 grade-school-level multiple-choice science questions from US standardized te...
-
ARIA, Accessible Rich Internet Applications, is the W3C standard published by the WAI that defines a set of HTML attributes — primarily role, aria-*, and state properties — for communicating the pu...
-
AutoGen is the multi-agent conversation framework released by Microsoft Research in 2023 (Wu et al.) that models agent interaction as natural-language conversation between agents, where each agent ...
-
AWQ, short for Activation-aware Weight Quantization, is a quantization technique introduced in a 2023 paper by Lin et al. (MIT) that produces 4-bit quantized LLMs with substantially better quality ...
-
Backpropagation, often shortened to backprop, is the algorithm that computes gradients of a neural network's loss with respect to its weights by applying the chain rule of calculus backward through...
-
Barcode recognition is the computer-vision technology that detects and decodes one-dimensional (linear) and two-dimensional (matrix) barcodes from images, video, or scanner inputs — the backbone of...
-
Batch size is the number of training examples processed together in one forward-backward pass before the optimizer updates the model weights, a fundamental hyperparameter affecting training speed, ...
-
Bayesian inference is the framework for updating beliefs about parameters or hypotheses as evidence accumulates, named for Thomas Bayes (1701-1761) whose theorem provides the mathematical machinery...
-
BEIR, short for Benchmarking IR (Information Retrieval), is a heterogeneous benchmark for evaluating retrieval systems introduced by Thakur et al. in 2021, covering 18 datasets across diverse domai...
-
BIG-bench, short for Beyond the Imitation Game Benchmark, is a collaborative LLM evaluation benchmark released in 2022 with contributions from 444 authors at 132 institutions. The benchmark contain...
-
BLEU, short for Bilingual Evaluation Understudy, is the classical machine translation evaluation metric introduced by Papineni et al. at IBM in 2002, scoring candidate translations against one or m...
-
BM25, short for Best Matching 25, is the classical lexical retrieval scoring function that has dominated information retrieval since Stephen Robertson and Karen Spärck Jones developed it at City Un...
-
BPE, Byte-Pair Encoding, is the subword tokenization algorithm originally invented for data compression by Philip Gage in 1994 and adapted for neural machine translation by Sennrich et al. in 2016,...
-
A canary release is the deployment pattern where a new version of a model or service is exposed to a small fraction of real production traffic first — typically 1-5% — while the bulk of traffic con...
-
Causal inference is the family of techniques for estimating cause-and-effect relationships from data — answering "did X cause Y?" rather than "are X and Y associated?" — and has emerged as one of t...
-
Chain-of-Thought, abbreviated CoT, is the prompting technique where an LLM is asked to explicitly reason step by step before producing a final answer, dramatically improving accuracy on math, logic...
-
Change Data Capture, abbreviated CDC, is the technique of identifying and propagating row-level changes (inserts, updates, deletes) from source databases to downstream systems in near-real-time, re...
-
Chunked prefill is an LLM serving optimization that splits the prefill phase (processing the input prompt) of long-context requests into smaller chunks, interleaving them with the decode phase (gen...
-
Chunking is the deceptively important preprocessing step in RAG where source documents are split into smaller passages that fit within an embedding model's context window and produce coherent retri...
-
CLIP, short for Contrastive Language-Image Pretraining, is the dual-encoder model architecture introduced by OpenAI in 2021 (Radford et al.) that learns a shared embedding space for images and text...
-
Cohort analysis is the analytical technique of grouping users (or customers, or any other unit) by a shared characteristic at a defined starting point — typically the month they first signed up, th...
-
ColBERT (Contextualized Late Interaction over BERT) is a late-interaction retrieval architecture introduced by Khattab and Zaharia at Stanford in 2020 that occupies a middle ground between bi-encod...
-
Collaborative Redline Review is an enterprise discipline within Version History & Redlining focused on governing how organizations create, review, route, approve, secure, publish, revise, archive,...
-
Color contrast ratio is the quantitative measure of luminance difference between two colors — typically text foreground and background — required by WCAG and accessibility laws to ensure readabilit...
-
Compliance Dashboard Reporting is an enterprise discipline within Compliance & Read Tracking focused on governing how organizations create, review, route, approve, secure, publish, revise, archive...
-
Compliance Escalation Reporting is an enterprise discipline within Compliance & Read Tracking focused on governing how organizations create, review, route, approve, secure, publish, revise, archiv...
-
Compliance Platform Migration is an enterprise discipline within Migration & Modernization focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, ...
-
Compliance Read Reporting is an enterprise discipline within Compliance & Read Tracking focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and...
-
Conditional Workflow Routing is an enterprise discipline within Workflow & Approvals focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and mo...
-
A confidence interval is the range of values around a sample estimate that is statistically likely to contain the true population parameter, providing a precision-aware alternative to (or accompani...
-
Constitutional AI, abbreviated CAI, is an alignment approach introduced by Anthropic in a 2022 paper that uses AI-generated critiques and revisions guided by a written set of principles (the "const...
-
A content filter is a rule-based or model-based system that blocks LLM outputs (or inputs) containing prohibited content categories — violence, sexual content, hate speech, self-harm, etc. — and is...
-
Content Platform Consolidation is an enterprise discipline within Migration & Modernization focused on governing how organizations create, review, route, approve, secure, publish, revise, archive,...
-
Content Repository Governance is an enterprise discipline within Document Management focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and mo...
-
The context window is the maximum number of tokens an LLM can process in a single forward pass — input prompt plus generated output — defining the upper bound on how much information the model can ...
-
Contextual Retrieval is a chunk-enrichment technique published by Anthropic in 2024 that prepends to each chunk a 50-100 token contextual summary explaining what the chunk is about and where it sit...
-
Continuous batching, sometimes called dynamic batching or in-flight batching, is an LLM serving technique where new incoming requests join a running batch immediately rather than waiting for the cu...
-
Contract Approval Governance is an enterprise discipline within Workflow & Approvals focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and mo...
-
Contract Lifecycle Management is an enterprise discipline within Document Management focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and mo...
-
Contract Repository Governance is an enterprise discipline within Document Management focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and m...
-
Contract System Migration is an enterprise discipline within Migration & Modernization focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and ...
-
Controlled Change Governance is an enterprise discipline within Version History & Redlining focused on governing how organizations create, review, route, approve, secure, publish, revise, archive,...
-
Controlled Document Distribution is an enterprise discipline within Document Management focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and...
-
Controlled Version Publishing is an enterprise discipline within Version History & Redlining focused on governing how organizations create, review, route, approve, secure, publish, revise, archive...
-
CrewAI is the role-based multi-agent framework released in late 2023 by João Moura that organizes agent collaboration around the metaphor of a crew — each agent has a defined role (Researcher, Writ...
-
A cross-encoder is a neural model architecture used for fine-grained relevance scoring in reranking , where the query and a candidate document are concatenated and fed into a single transformer tha...
-
Cross-Platform Document Migration is an enterprise discipline within Migration & Modernization focused on governing how organizations create, review, route, approve, secure, publish, revise, archi...
-
DALL-E is the proprietary text-to-image generation model family from OpenAI, named after Salvador Dalí and Pixar's WALL-E, released in three major generations (DALL-E January 2021, DALL-E 2 April 2...
-
Data aggregation is the process of consolidating data from multiple sources into a unified, queryable view, often with summarization or rollup applied to produce metrics, reports, or feature vector...
-
A data catalog is the centralized inventory of an organization's data assets, capturing for each dataset its schema, owner, description, classification, lineage, quality metrics, access policies, a...
-
A data lakehouse is the architectural pattern, formalized by Databricks researchers in 2020 (Armbrust et al.), that combines the open-format flexibility and low storage cost of a data lake with the...
-
Data lineage is the documented record of where data came from, how it has been transformed at every step, and where it currently flows — the complete provenance graph from source system to final co...
-
Data Loss Prevention, abbreviated DLP, is the family of policies and technologies that detect and block unauthorized exfiltration of sensitive data — by users, applications, or AI systems — across ...
-
Data mesh is a sociotechnical architecture for organizing data ownership, coined by Zhamak Dehghani at ThoughtWorks in 2019, where data products are owned by the business domains that produce them ...
-
Data mining is the broad discipline of extracting patterns, structure, and actionable information from large datasets — a field that predates the modern AI wave by decades and that now provides the...
-
Data parallelism is the simplest and most common form of distributed training, where the model is replicated on each GPU and different micro-batches of training data are processed simultaneously ac...
-
Data poisoning is the adversarial attack where malicious actors inject specially crafted records into a model's training data to degrade overall accuracy, introduce backdoors, or cause targeted mis...
-
Data provenance is the metadata that captures the origin and history of a piece of data, often used interchangeably with data lineage but more frequently referring to the per-record or per-dataset ...
-
Data residency is the requirement that data — including data used in AI training, inference, retrieval, and logging — remain within a specific geographic jurisdiction, typically driven by national ...
-
Data visualization is the discipline of representing data graphically to support understanding, communication, and decision-making — a field with foundations in William Playfair's 18th-century inve...
-
A data warehouse is the centralized analytical database designed for query-heavy reporting and analytics workloads, characterized by columnar storage, denormalized dimensional models, separation of...
-
A datasheet for a dataset is a structured documentation artifact for an ML training or evaluation dataset, proposed by Gebru et al. in a 2018 paper "Datasheets for Datasets" and modeled on the elec...
-
Deduplication is the process of identifying and consolidating duplicate or near-duplicate records in a dataset, a foundational step in any serious data pipeline and a make-or-break preprocessing st...
-
DeepSpeed is an open-source deep learning optimization library released by Microsoft Research in 2020 that provides memory-efficient training, distributed inference, and a suite of techniques for s...
-
Defensible disposition is the operational practice of destroying records at the end of their retention period in a manner that can withstand legal, regulatory, and audit scrutiny — proving that des...
-
Dense retrieval is the family of retrieval techniques where queries and documents are encoded as dense vectors in a continuous embedding space and matched by cosine similarity or dot product, in co...
-
Differential Privacy, abbreviated DP, is the rigorous mathematical framework for measuring and bounding privacy loss when releasing computations over sensitive data, formalized by Cynthia Dwork and...
-
Diffusion models are the generative model family that produces images, audio, and video by learning to reverse a gradual noising process — starting from pure Gaussian noise and iteratively denoisin...
-
Diffusion Transformer, abbreviated DiT, is the architectural pattern (Peebles and Xie, 2022) that replaces the U-Net backbone traditionally used in diffusion models with a Transformer, demonstratin...
-
Digital Approval Lifecycle is an enterprise discipline within Workflow & Approvals focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and moni...
-
Digital Compliance Monitoring is an enterprise discipline within Compliance & Read Tracking focused on governing how organizations create, review, route, approve, secure, publish, revise, archive,...
-
Digital File Governance is an enterprise discipline within Document Management focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and monitor ...
-
Digital Records Modernization is an enterprise discipline within Migration & Modernization focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, ...
-
Digital Redline Governance is an enterprise discipline within Version History & Redlining focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, a...
-
Document Accessibility Governance is an enterprise discipline within Document Management focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, an...
-
Document Accountability Governance is an enterprise discipline within Compliance & Read Tracking focused on governing how organizations create, review, route, approve, secure, publish, revise, arc...
-
Document Archival Strategy is an enterprise discipline within Document Management focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and monit...
-
Document Change Tracking is an enterprise discipline within Version History & Redlining focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and...
-
Document Classification is an enterprise discipline within Document Management focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and monitor ...
-
Document Compliance Tracking is an enterprise discipline within Document Management focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and mon...
-
Document Consumption Analytics is an enterprise discipline within Compliance & Read Tracking focused on governing how organizations create, review, route, approve, secure, publish, revise, archive...
-
Document Diff Analysis is an enterprise discipline within Version History & Redlining focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and m...
-
Document Governance Modernization is an enterprise discipline within Migration & Modernization focused on governing how organizations create, review, route, approve, secure, publish, revise, archi...
-
Document Intelligence is the umbrella discipline (and the specific product names from Azure, Google, AWS, and IBM) for AI-powered extraction of structured information from semi-structured documents...
-
Document Metadata Management is an enterprise discipline within Document Management focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and mon...
-
Document Repository Modernization is an enterprise discipline within Migration & Modernization focused on governing how organizations create, review, route, approve, secure, publish, revise, archi...
-
Document Restore Automation is an enterprise discipline within Version History & Redlining focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, ...
-
Document Review Workflow is an enterprise discipline within Workflow & Approvals focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and monito...
-
Document Revision History is an enterprise discipline within Version History & Redlining focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, an...
-
Document Revision Intelligence is an enterprise discipline within Version History & Redlining focused on governing how organizations create, review, route, approve, secure, publish, revise, archiv...
-
Document Security Governance is an enterprise discipline within Document Management focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and mon...
-
Document skew correction, also called deskewing, is the image-preprocessing step in document scanning that detects and corrects the rotation angle of a scanned page so subsequent OCR , OMR , and la...
-
Document Taxonomy Management is an enterprise discipline within Document Management focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and mon...
-
Document Version Governance is an enterprise discipline within Version History & Redlining focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, ...
-
Documentum Modernization is an enterprise discipline within Migration & Modernization focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and m...
-
DPO, short for Direct Preference Optimization, is an alignment technique introduced by Rafailov et al. in a May 2023 paper that achieves RLHF -quality results without training a separate reward mod...
-
A draft model is the small fast model used in speculative decoding to propose candidate tokens that the larger target model then verifies in parallel. The draft model must be from the same model fa...
-
Dynamic Workflow Routing is an enterprise discipline within Workflow & Approvals focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and monito...
-
ECM Migration Strategy is an enterprise discipline within Migration & Modernization focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and mon...
-
eDiscovery, electronic discovery, is the legal process of identifying, preserving, collecting, processing, reviewing, analyzing, and producing electronically stored information (ESI) in response to...
-
ELT, Extract-Load-Transform, is the modern inversion of the classical ETL Pipeline pattern: rather than transforming data in flight between source and destination, ELT loads raw data into the desti...
-
An embedding is a dense numerical vector representation of a piece of content — text, image, audio, or structured record — that encodes semantic meaning in a way amenable to similarity computation,...
-
Employee Acknowledgement Tracking is an enterprise discipline within Compliance & Read Tracking focused on governing how organizations create, review, route, approve, secure, publish, revise, arch...
-
Employee Policy Governance is an enterprise discipline within Compliance & Read Tracking focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, an...
-
Employee Read Tracking is an enterprise discipline within Compliance & Read Tracking focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and mo...
-
Encoder-decoder architecture is the original Transformer configuration from the 2017 "Attention Is All You Need" paper, where an encoder processes the input sequence into a contextualized represent...
-
Encoder-only models are the family of Transformer variants that consist solely of encoder layers (no decoder) and are optimized for understanding tasks — classification, named-entity recognition, s...
-
Enterprise Archive Consolidation is an enterprise discipline within Migration & Modernization focused on governing how organizations create, review, route, approve, secure, publish, revise, archiv...
-
Enterprise Attestation Reporting is an enterprise discipline within Compliance & Read Tracking focused on governing how organizations create, review, route, approve, secure, publish, revise, archi...
-
Enterprise Content Transformation is an enterprise discipline within Migration & Modernization focused on governing how organizations create, review, route, approve, secure, publish, revise, archi...
-
Enterprise Document Management is an enterprise discipline within Document Management focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and m...
-
Enterprise Knowledge Compliance is an enterprise discipline within Compliance & Read Tracking focused on governing how organizations create, review, route, approve, secure, publish, revise, archiv...
-
Enterprise Knowledge Migration is an enterprise discipline within Migration & Modernization focused on governing how organizations create, review, route, approve, secure, publish, revise, archive,...
-
Enterprise Process Automation is an enterprise discipline within Workflow & Approvals focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and m...
-
Enterprise Redline Auditing is an enterprise discipline within Version History & Redlining focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, ...
-
Enterprise Search Management is an enterprise discipline within Document Management focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and mon...
-
Entity linking is the natural-language-processing task of identifying mentions of entities in unstructured text and linking them to their canonical identifiers in a knowledge base — "Apple" in a ne...
-
Escalation Workflow Management is an enterprise discipline within Workflow & Approvals focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and ...
-
An ETL pipeline (Extract, Transform, Load) is the orchestrated workflow that pulls data from source systems, applies transformations to clean, conform, and enrich it, and writes the result to a des...
-
The EU AI Act is the European Union's comprehensive horizontal regulation of artificial intelligence, formally adopted in 2024 and rolling into effect across staged compliance deadlines from 2024 t...
-
An evaluation card is a structured documentation artifact for an AI model evaluation, describing the benchmark used, the evaluation methodology, the prompts and conditions, the results, and the lim...
-
Exception Routing Workflows is an enterprise discipline within Workflow & Approvals focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and mon...
-
Exploratory Data Analysis, abbreviated EDA, is the open-ended phase of a data-science workflow where the analyst examines a dataset visually and statistically to discover its structure, distributio...
-
Feature engineering is the data-science discipline of transforming raw data into the predictor variables (features) that a model actually consumes — a craft that often determines model performance ...
-
A feature store is the centralized infrastructure that computes, stores, and serves the features used by machine-learning models — both during training (where features must match the model's expect...
-
Federated Learning is the distributed machine-learning paradigm where multiple clients (mobile devices, hospitals, banks, edge nodes) collaboratively train a shared model without sharing their raw ...
-
The feed-forward network, abbreviated FFN, is the second sublayer in each Transformer block (the first being multi-head attention ), responsible for applying nonlinear transformations to each posit...
-
Few-shot prompting is the technique of including 2-10 worked examples of input-output pairs in the prompt before the actual task input, letting the LLM learn the desired behavior from the examples ...
-
FlashAttention is an exact attention algorithm introduced by Tri Dao et al. in a 2022 paper that dramatically accelerates self-attention by tiling computations to keep intermediate tensors in fast ...
-
Form field detection is the document-AI capability that locates and classifies fillable regions on a structured form — text fields, checkboxes, radio buttons, dropdown selections, signature lines, ...
-
FSDP, short for Fully Sharded Data Parallel, is a distributed training technique built into PyTorch since version 1.11 (2022) that shards model parameters, gradients, and optimizer states across mu...
-
Full fine-tuning is the training approach that updates all of a pretrained model's weights on task-specific data, in contrast to PEFT methods like LoRA , adapter layers, and prefix tuning that upda...
-
Function calling is the structured tool-use capability built into modern LLMs where the model emits a JSON object specifying a function name and arguments rather than free-form text, enabling relia...
-
Fuzzy matching is the family of techniques for identifying records that refer to the same entity despite differing in spelling, formatting, abbreviations, transposition, or capitalization — the pra...
-
GGUF, short for GPT-Generated Unified Format, is the binary file format used by Llama.cpp and Ollama to store quantized LLM weights, metadata, and tokenizer configuration in a single self-contained...
-
Governed Content Consolidation is an enterprise discipline within Migration & Modernization focused on governing how organizations create, review, route, approve, secure, publish, revise, archive,...
-
Governed Read Receipts is an enterprise discipline within Compliance & Read Tracking focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and mo...
-
Governed SharePoint Replacement is an enterprise discipline within Migration & Modernization focused on governing how organizations create, review, route, approve, secure, publish, revise, archive...
-
Governed Task Automation is an enterprise discipline within Workflow & Approvals focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and monito...
-
Governed Workflow Automation is an enterprise discipline within Workflow & Approvals focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and mo...
-
GPTQ is a one-shot quantization technique introduced by Frantar et al. in 2022 that produces 3-bit or 4-bit quantized LLMs with minimal accuracy loss using approximate second-order information (an ...
-
Gradient accumulation is a training technique that simulates larger effective batch sizes by accumulating gradients over multiple forward-backward passes before applying an optimizer step, allowing...
-
Gradient checkpointing is a memory-saving training technique that trades compute for memory by recomputing intermediate activations during the backward pass rather than storing them from the forwar...
-
Gradient descent is the iterative optimization algorithm that updates a neural network's weights by stepping in the direction opposite to the gradient of the loss function, gradually moving toward ...
-
GraphRAG is the retrieval-augmented generation pattern published by Microsoft Research in 2024 (Edge et al. "From Local to Global: A Graph RAG Approach to Query-Focused Summarization") that uses a ...
-
Grouped-Query Attention, abbreviated GQA, is a multi-head attention variant introduced by Ainslie et al. in a 2023 Google paper that shares key and value projections across groups of query heads, r...
-
GSM8K, short for Grade School Math 8K, is a benchmark introduced by OpenAI in 2021 containing 8,500 grade-school-level multi-step math word problems requiring 2-8 steps of reasoning to solve. The b...
-
Guardrails are programmable policy enforcement layers around LLM applications that validate inputs and outputs against rule-based, regex, classifier, or LLM-judged criteria. Unlike model-level refu...
-
HELM, short for Holistic Evaluation of Language Models, is a benchmark framework introduced by Stanford's Center for Research on Foundation Models (CRFM) in 2022 that evaluates LLMs across 7 metric...
-
Helpful, Harmless, and Honest, often abbreviated HHH, is the canonical three-part value framework introduced by Anthropic in a 2021 paper ("A General Language Assistant as a Laboratory for Alignmen...
-
Historical Version Auditing is an enterprise discipline within Version History & Redlining focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, ...
-
Homomorphic Encryption, abbreviated HE, is the family of cryptographic schemes that allow computations on encrypted data to produce encrypted results that, when decrypted, match the result of opera...
-
HumanEval is a code-generation benchmark introduced by OpenAI alongside Codex in a 2021 paper, containing 164 hand-written Python programming problems with unit tests for automatic verification. Ea...
-
Hybrid search combines lexical retrieval (keyword-based, typically BM25 ) with semantic retrieval (vector-based dense retrieval ) and fuses the results, capturing both the precision of exact-match ...
-
HyDE, short for Hypothetical Document Embeddings, is a 2022 retrieval technique published by Gao et al. (CMU and Allen Institute) that improves zero-shot retrieval by having an LLM generate a hypot...
-
Hyland OnBase Migration is an enterprise discipline within Migration & Modernization focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and mo...
-
Hypothesis testing is the formal statistical procedure, developed by Fisher, Neyman, and Pearson in the early twentieth century, for evaluating whether observed data provides sufficient evidence to...
-
ICR, Intelligent Character Recognition, is the specialized variant of optical text recognition focused on handprinted and constrained handwritten characters — the technology that reads handwritten ...
-
Image binarization is the image-processing step that converts a grayscale or color document image into a strictly black-and-white (1-bit) image, separating foreground text and graphics from backgro...
-
In-Context Learning, abbreviated ICL, is the emergent ability of large LLMs to learn new tasks from examples provided in the prompt at inference time, without any weight updates — discovered as a p...
-
Indirect prompt injection is a specific class of prompt injection attacks where the malicious instructions are embedded in third-party content the LLM processes — web pages it browses, documents in...
-
Information governance, abbreviated IG, is the comprehensive framework by which an organization manages all of its information — structured and unstructured, internal and external, digital and phys...
-
Instruction tuning is the post-pretraining adaptation technique that teaches a base LLM to follow natural-language instructions by training it on datasets of (instruction, response) pairs. The tech...
-
ISO/IEC 42001 is the international standard for AI Management Systems (AIMS), published in December 2023 by the joint ISO/IEC technical committee, and the first ISO standard specifically certifying...
-
A jailbreak is an attack on an LLM that bypasses the model's safety training to elicit responses the model was trained to refuse — typically harmful instructions, restricted content, or operations ...
-
JSON mode is an LLM output constraint that forces the model to produce valid JSON, eliminating the parsing failures that plague free-text JSON generation. OpenAI introduced JSON mode in November 20...
-
Keyboard navigation is the accessibility principle and engineering practice ensuring that every interactive element of a digital experience can be reached and operated using only the keyboard, with...
-
Knowledge distillation is the model-compression technique where a smaller "student" model learns to mimic the behavior of a larger "teacher" model, typically by training the student on the teacher'...
-
Knowledge Distribution Governance is an enterprise discipline within Compliance & Read Tracking focused on governing how organizations create, review, route, approve, secure, publish, revise, arch...
-
A knowledge graph is the structured representation of entities and the relationships between them as a graph — nodes for entities (people, organizations, products, places, concepts), edges for rela...
-
Knowledge Repository Management is an enterprise discipline within Document Management focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and ...
-
KTO, short for Kahneman-Tversky Optimization, is an alignment technique introduced by ContextualAI in 2024 that draws on prospect theory from behavioral economics to align LLMs using unpaired good ...
-
LangGraph is the agent-orchestration framework released by LangChain in early 2024 that models agentic workflows as directed graphs of nodes (each node is a function or LLM call) and edges (transit...
-
Layer normalization, often abbreviated LayerNorm, is a normalization technique introduced by Ba, Kiros, and Hinton in 2016 that normalizes activations across the feature dimension within each token...
-
Layout analysis is the document AI capability that identifies and labels structural regions within a document image — titles, headings, paragraphs, tables, figures, lists, headers, footers, page nu...
-
Learning rate is the hyperparameter that controls how large a step the optimizer takes in the direction of the negative gradient during training — too small and training is slow, too large and trai...
-
Legacy Archive Migration is an enterprise discipline within Migration & Modernization focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and m...
-
Legacy ECM Transformation is an enterprise discipline within Migration & Modernization focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and ...
-
Legacy Workflow Modernization is an enterprise discipline within Migration & Modernization focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, ...
-
Legal Document Governance is an enterprise discipline within Document Management focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and monito...
-
A legal hold, also called a litigation hold or preservation hold, is the formal directive issued by an organization (typically through its General Counsel) to suspend the normal disposition of reco...
-
Levenshtein distance, also called edit distance, is the minimum number of single-character insertions, deletions, or substitutions required to transform one string into another — a quantification o...
-
Live Compliance Reporting is an enterprise discipline within Compliance & Read Tracking focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and...
-
Llama.cpp is an open-source LLM inference engine written in pure C/C++ by Georgi Gerganov, released in March 2023, that enables CPU and consumer-GPU inference of quantized LLMs with minimal depende...
-
Locality-Sensitive Hashing, abbreviated LSH, is the family of hashing techniques designed so that similar inputs collide into the same hash bucket with high probability while dissimilar inputs go t...
-
LoRA, short for Low-Rank Adaptation, is a parameter-efficient fine-tuning technique introduced by Microsoft Research in a 2021 paper by Hu et al. that has become the dominant approach to adapting l...
-
LoRA rank, often denoted r, is the dimensionality of the low-rank decomposition that LoRA adapters use to approximate weight updates — a key hyperparameter that controls the trade-off between adapt...
-
Mamba is the selective-state-space-model architecture introduced by Albert Gu (CMU) and Tri Dao (Princeton) in December 2023, the breakthrough State Space Model that demonstrated competitive langua...
-
Mandatory Document Distribution is an enterprise discipline within Compliance & Read Tracking focused on governing how organizations create, review, route, approve, secure, publish, revise, archiv...
-
Master Data Management, abbreviated MDM, is the discipline of creating and maintaining a single authoritative record (the "golden record" or "master") for each business entity — customer, product, ...
-
Maximal Marginal Relevance, abbreviated MMR, is a retrieval reranking algorithm published by Carbonell and Goldstein in 1998 that balances relevance against diversity, ensuring that the top-k retur...
-
Megatron-LM is an open-source LLM training framework developed by NVIDIA, originally introduced in 2019 and expanded with the Megatron-Turing NLG 530B collaboration with Microsoft. The framework pr...
-
A membership inference attack, abbreviated MIA, is the privacy threat where an adversary determines whether a specific data record was in a model's training set by observing the model's behavior — ...
-
Meta-prompting is the family of techniques where an LLM is used to generate, evaluate, or improve prompts that are then used by the same or another LLM for downstream tasks — essentially using AI t...
-
MICR, Magnetic Ink Character Recognition, is the specialized character-recognition technology developed for the banking industry in the 1950s by Stanford Research Institute and General Electric to ...
-
MinHash is the locality-sensitive hashing technique invented by Andrei Broder at AltaVista in 1997 to detect near-duplicate web pages, and remains the workhorse of large-scale near-duplicate detect...
-
Mixed precision training is a technique that uses lower-precision floating point (typically FP16 or BF16) for most computations while keeping a master copy of weights in full FP32 precision, dramat...
-
Mixture of Experts, abbreviated MoE, is the neural-network architecture in which different inputs are routed to different specialized subnetworks (the "experts") via a learned gating mechanism, dra...
-
MMLU, short for Massive Multitask Language Understanding, is a benchmark introduced by Hendrycks et al. in 2020 that tests LLM knowledge and reasoning across 57 subjects ranging from elementary mat...
-
A model card is a structured documentation artifact for an AI model covering its intended uses, training data, evaluation results, limitations, and ethical considerations, proposed by Mitchell et a...
-
Model drift is the degradation of a deployed model's performance over time, driven by changes in the data the model encounters (data drift) or in the relationships between inputs and the desired ou...
-
A model extraction attack is the security threat where an adversary queries a deployed AI model repeatedly to reverse-engineer its parameters, decision boundaries, or training data — effectively st...
-
A model inversion attack is the privacy threat where an adversary exploits a deployed model's behavior to reconstruct features of training data — recovering faces from face-recognition models, reco...
-
Model monitoring is the continuous observation of a deployed model's behavior, quality, and operational health in production — capturing the metrics, alerts, dashboards, and traces that let an orga...
-
Model pruning is the family of model-compression techniques that remove weights, neurons, attention heads, or entire layers from a trained neural network to reduce its size and inference cost while...
-
A model registry is the centralized catalog of trained machine-learning models — including LLMs , LoRA adapters, embedding models, classifiers, and rerankers — that tracks versions, lineage, metric...
-
MRKL, short for Modular Reasoning, Knowledge and Language (pronounced "miracle"), is an agent architecture introduced by AI21 Labs in 2022 that combines an LLM with external expert modules for task...
-
MT-Bench is a benchmark for evaluating chat-tuned LLMs on multi-turn conversational tasks, introduced by LMSYS in 2023 alongside the Chatbot Arena leaderboard. The benchmark contains 80 multi-turn ...
-
MTEB, short for Massive Text Embedding Benchmark, is the standard benchmark for evaluating embedding models , introduced by Hugging Face and Cohere researchers in 2023. The benchmark covers 8 task ...
-
A multi-agent system, abbreviated MAS in the AI literature, is the architectural pattern where multiple LLM -powered agents — each with specialized roles, tools, and capabilities — collaborate to s...
-
Multi-head attention is the parallel-attention scheme in the Transformer that runs multiple self-attention operations in parallel with different linear projections of the input, then concatenates t...
-
A multimodal LLM is a large language model that natively accepts inputs in multiple modalities — text, images, audio, video, sometimes documents and structured data — and reasons across them in a s...
-
A multimodal Transformer is a Transformer architecture explicitly designed to process multiple input modalities — text, image, audio, video, structured data — through a unified attention mechanism,...
-
Multi-Query Attention, abbreviated MQA, is an extreme variant of Grouped-Query Attention introduced by Shazeer in a 2019 paper, where all query heads share a single key and value projection. MQA re...
-
Multi-Step Workflow Automation is an enterprise discipline within Workflow & Approvals focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and ...
-
Named Entity Recognition, abbreviated NER, is the natural-language-processing task of identifying and classifying spans of text as entities of specific types — typically person names, organizations...
-
The NIST AI Risk Management Framework, abbreviated AI RMF and published as NIST AI 100-1 in January 2023 with a Generative AI Profile (AI 600-1) added in July 2024, is the United States federal vol...
-
OCR, Optical Character Recognition, is the family of computer vision techniques that extract machine-readable text from images of printed, typewritten, or handwritten content — the foundational ing...
-
Ollama is an open-source LLM serving wrapper around Llama.cpp that adds a clean REST API, a Docker-Hub-style model registry, and a one-line install experience, making local LLM inference accessible...
-
OMR, Optical Mark Recognition, is the specialized scanning technology that detects the presence or absence of marks (bubbles, checkboxes, X marks) in predefined regions on a form, enabling rapid di...
-
An ontology in the AI and knowledge-engineering sense is the formal, structured description of the concepts, relationships, and constraints in a domain — defining what classes of entities exist, wh...
-
ORPO, short for Odds Ratio Preference Optimization, is an alignment technique introduced by Hong et al. in early 2024 that merges SFT and preference optimization into a single training stage, elimi...
-
Outlier detection, also called anomaly detection, is the family of techniques for identifying data points that deviate substantially from the rest of a dataset — points that may indicate data-quali...
-
PagedAttention is the memory management algorithm at the heart of vLLM , introduced in a 2023 paper by UC Berkeley researchers that solved one of the most painful problems in LLM serving: KV cache ...
-
Parallel Approval Workflows is an enterprise discipline within Workflow & Approvals focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and mon...
-
PDF/UA, PDF Universal Accessibility, is the ISO standard (ISO 14289-1:2014, revised 2024) for accessible PDF documents — the PDF-specific complement to WCAG that governs how PDFs must be tagged, st...
-
PEFT, short for Parameter-Efficient Fine-Tuning, is the umbrella term for a family of techniques that adapt large pretrained models by training only a tiny fraction of their parameters — typically ...
-
Perplexity is the exponential of the average negative log-likelihood that a language model assigns to a held-out text corpus, with lower perplexity indicating the model finds the text more probable...
-
Phonetic matching is the family of algorithms that encode strings (typically names) by how they sound rather than how they are spelled, allowing "Smith" and "Smyth" or "Catherine" and "Katharine" t...
-
PII redaction is the process of automatically detecting and removing or masking personally identifiable information — names, email addresses, phone numbers, Social Security numbers, dates of birth,...
-
Pipeline parallelism is a distributed training technique that partitions a neural network's layers across multiple GPUs or nodes, with each device handling a contiguous slice of the model. During t...
-
Plan-and-Execute is an agentic pattern where an LLM first generates a complete plan as a list of steps, then executes each step in sequence — separating the planning phase from execution rather tha...
-
Policy Approval Automation is an enterprise discipline within Workflow & Approvals focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and moni...
-
Policy Attestation Governance is an enterprise discipline within Compliance & Read Tracking focused on governing how organizations create, review, route, approve, secure, publish, revise, archive,...
-
Policy Document Governance is an enterprise discipline within Document Management focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and monit...
-
Policy Lifecycle Automation is an enterprise discipline within Document Management focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and moni...
-
Policy Read Auditing is an enterprise discipline within Compliance & Read Tracking focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and moni...
-
Policy Repository Transformation is an enterprise discipline within Migration & Modernization focused on governing how organizations create, review, route, approve, secure, publish, revise, archiv...
-
Positional encoding is the mechanism that gives Transformer models information about the order of tokens in a sequence, since self-attention by itself is permutation-equivariant and cannot distingu...
-
Prefix caching is an LLM inference optimization that reuses the computed KV cache for shared prompt prefixes across multiple requests, eliminating redundant computation when many requests share the...
-
Prefix tuning is a PEFT technique introduced by Li and Liang (2021) that prepends a small sequence of learned continuous vectors (the prefix) to every layer's attention input, allowing task adaptat...
-
Prompt chaining is the architectural pattern where a complex task is decomposed into a sequence of simpler LLM calls, with the output of each call feeding the next, rather than asking one model to ...
-
Prompt injection is a class of attacks where an attacker inserts instructions into LLM input that override or subvert the application's intended behavior, named by analogy to SQL injection. Direct ...
-
Prompt tuning, also called soft prompt tuning, is a PEFT technique introduced by Lester, Al-Rfou, and Constant (2021) that prepends a short sequence of learned continuous vectors directly to the in...
-
A p-value is the probability of observing data at least as extreme as the actual result, assuming the null hypothesis is true — the central output of frequentist hypothesis testing and one of the m...
-
QLoRA, short for Quantized Low-Rank Adaptation, is an extension of LoRA introduced by Dettmers et al. in 2023 that combines 4-bit quantization of the base model with LoRA adapter training, enabling...
-
Retrieval-Augmented Generation, universally abbreviated RAG, is the architectural pattern in which a large language model answers a question by first retrieving relevant passages from an external c...
-
RAGAS, short for Retrieval-Augmented Generation Assessment, is an evaluation framework specifically designed for RAG pipelines, released as an open-source library in 2023 and adopted by many enterp...
-
RDF, the Resource Description Framework, is the W3C-standardized data model for representing information on the web as a graph of triples — subject-predicate-object — that has been the substrate of...
-
ReAct, short for Reasoning and Acting, is an agentic LLM framework introduced by Yao et al. in a 2022 Google paper that interleaves reasoning traces with action calls, producing more reliable agent...
-
Read Compliance Automation is an enterprise discipline within Compliance & Read Tracking focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, an...
-
Record linkage, also called entity matching or entity resolution, is the systematic process of identifying records across one or more databases that refer to the same real-world entity — a discipli...
-
Records management is the systematic discipline of identifying, classifying, storing, securing, retrieving, and ultimately disposing of an organization's records — the documentary evidence of busin...
-
Records Retention Management is an enterprise discipline within Document Management focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and mon...
-
A records series is a logical grouping of records that share common characteristics — same business function, same regulatory authority, same retention requirement, same disposition action — and th...
-
Red teaming for AI is the practice of having dedicated adversarial testers attempt to elicit harmful, biased, false, or otherwise problematic outputs from an LLM before deployment, modeled on red-t...
-
Redline Comparison Governance is an enterprise discipline within Version History & Redlining focused on governing how organizations create, review, route, approve, secure, publish, revise, archive...
-
Reflexion is an agent pattern introduced by Shinn et al. in a 2023 paper that adds explicit self-reflection and learning from past attempts, enabling agents to improve performance on repeated trial...
-
Refusal training is the post-training technique that teaches LLMs to decline requests for harmful, dangerous, or policy-violating content — a fundamental component of every commercial LLM 's safety...
-
Regression analysis is the family of statistical techniques for modeling the relationship between a dependent variable and one or more independent (predictor) variables, with origins in Francis Gal...
-
Regulated Document Control is an enterprise discipline within Document Management focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and monit...
-
Regulatory Distribution Tracking is an enterprise discipline within Compliance & Read Tracking focused on governing how organizations create, review, route, approve, secure, publish, revise, archi...
-
Reranking is the second-pass retrieval step in production RAG pipelines where an initial set of candidates from dense retrieval or hybrid search is re-scored by a more accurate but slower model, ty...
-
Residual connections, also called skip connections, are direct paths from a layer's input to its output that bypass the intermediate computation, introduced by He et al. in the 2015 ResNet paper an...
-
A retention schedule is the official, organization-wide policy document that specifies how long each category of records must be kept, what triggers the start of the retention clock, and what dispo...
-
Reverse ETL is the operational pattern, named and popularized around 2020-2021, where data is synchronized from the analytical data warehouse back into operational SaaS systems — pushing customer s...
-
The right to explanation is the legal and ethical principle that individuals affected by automated decisions are entitled to receive a meaningful explanation of how those decisions were made, inclu...
-
RLHF, short for Reinforcement Learning from Human Feedback, is the alignment technique introduced by OpenAI in InstructGPT (2022) and used to train ChatGPT, Claude, Gemini, and most major commercia...
-
RMSNorm, short for Root Mean Square Layer Normalization, is a simplified layer normalization variant introduced by Zhang and Sennrich in a 2019 paper that omits the mean-centering step and bias par...
-
Role prompting is the technique of assigning the LLM an explicit role or persona in the prompt — "You are an expert tax attorney specializing in international transfer pricing" — to shift its tone,...
-
Rollback Recovery Governance is an enterprise discipline within Version History & Redlining focused on governing how organizations create, review, route, approve, secure, publish, revise, archive,...
-
RoPE, short for Rotary Position Embedding, is a positional encoding technique introduced by Su et al. in a 2021 paper that encodes absolute position via rotation matrices applied to query and key v...
-
ROUGE, short for Recall-Oriented Understudy for Gisting Evaluation, is a family of metrics for evaluating automatic summarization introduced by Chin-Yew Lin in 2004. The most-reported variants are ...
-
A safety classifier is a smaller specialized model that screens LLM inputs and outputs for harmful, toxic, or policy-violating content, typically deployed as a pre- or post-processing layer around ...
-
Schema drift is the unannounced change in the structure of a data source — a renamed column, a new field, a changed type, a removed table — that breaks downstream pipelines because consuming system...
-
Screen reader compatibility is the practical measure of whether a digital experience can be successfully consumed by users of screen-reading assistive technology — software that converts on-screen ...
-
Section 508 is the United States federal law, codified at 29 U.S.C. § 794d as part of the Rehabilitation Act of 1973 (amended substantially in 1998), that requires federal agencies to procure, deve...
-
Secure Multi-Party Computation, abbreviated SMPC or MPC, is the cryptographic technique that allows multiple parties to jointly compute a function over their private inputs without revealing those ...
-
Self-Ask is an agentic prompting pattern introduced by Press et al. in a 2022 paper that improves multi-hop question-answering by having the LLM explicitly decompose complex questions into simpler ...
-
Self-attention is the core mechanism of the Transformer architecture, allowing each position in a sequence to attend to every other position when computing its representation. The mechanism compute...
-
Self-consistency is a prompting technique published by Wang et al. (Google) in 2022 that improves chain-of-thought reasoning by sampling multiple independent reasoning chains from the LLM for the s...
-
Semantic search is the umbrella term for search systems that retrieve results based on meaning rather than literal keyword overlap, typically powered by dense retrieval over embedding vectors. The ...
-
SentencePiece is the language-independent subword tokenizer library released open-source by Google in 2018 (Kudo and Richardson), notable for treating the input as a raw Unicode stream without lang...
-
Sequential Approval Governance is an enterprise discipline within Workflow & Approvals focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and ...
-
SFT, short for Supervised Fine-Tuning, is the standard training phase that adapts a base LLM to follow instructions by training it on labeled examples of (input, desired output) pairs. SFT is typic...
-
Shadow deployment, also called shadow mode or shadow testing, is the technique of running a new model in production alongside the current model — receiving the same real-traffic inputs — but discar...
-
SharePoint Migration Governance is an enterprise discipline within Migration & Modernization focused on governing how organizations create, review, route, approve, secure, publish, revise, archive...
-
Slowly Changing Dimensions, abbreviated SCD, is the family of techniques for handling changes to dimension table attributes in a data warehouse over time, a problem Ralph Kimball formalized into si...
-
SPARQL (recursive acronym: SPARQL Protocol and RDF Query Language) is the W3C-standardized query language for RDF data, analogous to SQL for relational databases but designed for graph patterns rat...
-
Sparse attention is a family of self-attention variants that compute attention over only a structured subset of query-key pairs rather than the full quadratic set, dramatically reducing compute and...
-
Sparse retrieval is the umbrella term for retrieval methods that represent documents as high-dimensional sparse vectors (mostly zeros, with non-zero values only for the small number of distinct ter...
-
Special tokens are reserved entries in an LLM 's vocabulary that signal structural meaning rather than content — beginning of sequence, end of sequence, padding, separators between turns, system vs...
-
Speculative decoding is an LLM inference acceleration technique introduced by Google researchers in 2022 and refined by DeepMind in 2023 that uses a small fast "draft model" to propose multiple can...
-
Stable Diffusion is the open-weight latent diffusion model family released by Stability AI starting in August 2022, the model that brought high-quality text-to-image generation out of the cloud and...
-
A star schema is the dimensional data-modeling pattern, popularized by Ralph Kimball in The Data Warehouse Toolkit (1996), where a central fact table containing measurable events (sales transaction...
-
State Space Models, abbreviated SSMs in the modern AI context, are the family of sequence-modeling architectures inspired by classical control-theory state space representations, offering a credibl...
-
Statistical significance is the technical determination that an observed result is unlikely to have occurred by random chance alone if the null hypothesis were true — formalized through a p-value f...
-
Structured output is the broader class of LLM output constraints that force responses to conform to a specified schema — typically JSON Schema, a Pydantic class, a Zod schema, or a regex pattern — ...
-
Subword tokenization is the family of tokenization approaches that split text into units smaller than words but larger than characters — capturing meaningful morphological structure while ensuring ...
-
Survivorship rules are the policies in master data management and deduplication that determine which specific value should "survive" — be chosen as the golden record's canonical value — when duplic...
-
SwiGLU, short for Swish-Gated Linear Unit, is a feed-forward network variant introduced by Shazeer in a 2020 paper that combines the Swish activation function with a Gated Linear Unit (GLU) structu...
-
A system card is an extension of the model card concept that documents an end-to-end AI system — including the model, safety layers, deployment context, and known risks — rather than just the under...
-
A system prompt is the highest-priority instruction passed to an LLM at the start of a conversation, distinct from user messages, that defines the model's persona, capabilities, tone, constraints, ...
-
Tensor parallelism is a distributed training technique that splits individual layer computations across multiple GPUs, typically within a single node where high-bandwidth interconnects like NVLink ...
-
TensorRT-LLM is NVIDIA's open-source LLM inference framework, released in late 2023, that compiles transformer models into highly optimized CUDA kernels for the lowest possible latency on NVIDIA GP...
-
Tiktoken is OpenAI's open-source Rust-implemented BPE tokenizer library, released in late 2022, that powers the production tokenization for all OpenAI models and is widely adopted as a third-party ...
-
Time series analysis is the specialized branch of statistics and data science focused on data points indexed in time order — daily sales, hourly web traffic, monthly subscriber counts, second-by-se...
-
Tokenization is the preprocessing step that converts raw text into the integer token IDs that an LLM actually consumes, and the inverse step that converts model output IDs back to text — the bounda...
-
Tool use protocol refers to the formal interface by which an LLM agent invokes external tools, including the schema for declaring available tools, the format for invoking them, and the convention f...
-
Training Document Governance is an enterprise discipline within Compliance & Read Tracking focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, ...
-
The Transformer is the neural network architecture introduced in the seminal 2017 paper "Attention Is All You Need" by Vaswani et al. at Google, replacing recurrent networks (LSTMs, GRUs) as the fo...
-
Tree of Thoughts, abbreviated ToT, is a generalization of chain-of-thought introduced by Yao et al. (Princeton/Google DeepMind) in 2023 where the LLM generates multiple candidate next steps at each...
-
Triton Inference Server is NVIDIA's open-source model-serving framework, originally released in 2018, that serves any AI model — LLMs , vision, audio, classical ML — through a unified HTTP/gRPC API...
-
TruthfulQA is a benchmark introduced by Lin, Hilton, and Evans in 2021 that tests whether LLMs avoid generating false answers to questions designed to elicit common misconceptions, conspiracy theor...
-
A vector database is a specialized data store optimized for high-dimensional vector similarity search, the workhorse infrastructure behind RAG , semantic search, recommendation systems, and agentic...
-
Version Chain Management is an enterprise discipline within Version History & Redlining focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and...
-
Version Compliance Reporting is an enterprise discipline within Version History & Redlining focused on governing how organizations create, review, route, approve, secure, publish, revise, archive,...
-
Version Integrity Management is an enterprise discipline within Version History & Redlining focused on governing how organizations create, review, route, approve, secure, publish, revise, archive,...
-
Version Lifecycle Governance is an enterprise discipline within Version History & Redlining focused on governing how organizations create, review, route, approve, secure, publish, revise, archive,...
-
Version Rollback Management is an enterprise discipline within Version History & Redlining focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, ...
-
Version Traceability Governance is an enterprise discipline within Version History & Redlining focused on governing how organizations create, review, route, approve, secure, publish, revise, archi...
-
Vision Transformer, abbreviated ViT, is the architectural breakthrough published by Dosovitskiy et al. at Google in 2020 that adapted the Transformer architecture from natural language to computer ...
-
vLLM is an open-source LLM inference engine released by UC Berkeley researchers in 2023 that has become the dominant high-throughput serving framework for self-hosted LLM deployments. The framework...
-
The vocabulary of an LLM is the fixed set of tokens the model can encode and emit — typically 32,000 to 200,000 entries — determined by the tokenization algorithm trained on the model's corpus. Voc...
-
WCAG, the Web Content Accessibility Guidelines, is the international standard for web accessibility published by the W3C Web Accessibility Initiative (WAI), governing how digital content must be de...
-
Whisper is the automatic speech recognition (ASR) model family released open-source by OpenAI in September 2022, trained on 680,000 hours of multilingual and multitask supervised audio data and cap...
-
Workflow Analytics is an enterprise discipline within Workflow & Approvals focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and monitor crit...
-
Workflow Audit Management is an enterprise discipline within Workflow & Approvals focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and monit...
-
Workflow Compliance Monitoring is an enterprise discipline within Workflow & Approvals focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and ...
-
Workflow Exception Reporting is an enterprise discipline within Workflow & Approvals focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and mo...
-
Workflow Intelligence Reporting is an enterprise discipline within Workflow & Approvals focused on governing how organizations create, review, route, approve, secure, publish, revise, archive, and...
-
ZeRO, short for Zero Redundancy Optimizer, is a memory optimization technique introduced by Microsoft Research in 2019 that shards optimizer states, gradients, and (optionally) model parameters acr...
-
Zero-shot prompting is the technique of asking an LLM to perform a task by describing the task in natural language without providing any examples of the desired input-output behavior, relying entir...
-
Adam, short for Adaptive Moment Estimation, is the adaptive optimization algorithm introduced by Kingma and Ba in 2014 that has become the default optimizer for deep learning including LLM pretrain...
-
AdamW is a variant of the Adam optimizer introduced by Loshchilov and Hutter in 2017 that decouples weight decay from the gradient-based parameter updates, producing better generalization and train...
-
Adapter layers are a PEFT technique introduced by Houlsby et al. in 2019 that inserts small bottleneck feed-forward modules between the frozen layers of a pretrained transformer , training only the...
-
Agent memory is the persistent state that an agentic LLM system maintains across interactions, enabling continuity, learning, and personalization beyond a single conversation. Common memory types i...
-
ALiBi, short for Attention with Linear Biases, is a positional encoding technique introduced by Press, Smith, and Lewis in a 2021 paper that adds linear-distance penalties directly to attention sco...
-
Alignment tax is the term used to describe the trade-off between safety properties and raw capability that often emerges when LLMs undergo RLHF , refusal training , Constitutional AI , or other saf...
-
AlpacaEval is an automated evaluation framework for chat-tuned LLMs released by Stanford's Tatsu Lab in 2023, scoring models by win rate against a reference model (originally text-davinci-003) on a...
-
ARC, short for the AI2 Reasoning Challenge, is a benchmark introduced by Allen Institute for AI in 2018 containing 7,787 grade-school-level multiple-choice science questions from US standardized te...
-
AWQ, short for Activation-aware Weight Quantization, is a quantization technique introduced in a 2023 paper by Lin et al. (MIT) that produces 4-bit quantized LLMs with substantially better quality ...
-
Backpropagation, often shortened to backprop, is the algorithm that computes gradients of a neural network's loss with respect to its weights by applying the chain rule of calculus backward through...
-
Batch size is the number of training examples processed together in one forward-backward pass before the optimizer updates the model weights, a fundamental hyperparameter affecting training speed, ...
-
BEIR, short for Benchmarking IR (Information Retrieval), is a heterogeneous benchmark for evaluating retrieval systems introduced by Thakur et al. in 2021, covering 18 datasets across diverse domai...
-
BIG-bench, short for Beyond the Imitation Game Benchmark, is a collaborative LLM evaluation benchmark released in 2022 with contributions from 444 authors at 132 institutions. The benchmark contain...
-
BLEU, short for Bilingual Evaluation Understudy, is the classical machine translation evaluation metric introduced by Papineni et al. at IBM in 2002, scoring candidate translations against one or m...
-
Chunked prefill is an LLM serving optimization that splits the prefill phase (processing the input prompt) of long-context requests into smaller chunks, interleaving them with the decode phase (gen...
-
Constitutional AI, abbreviated CAI, is an alignment approach introduced by Anthropic in a 2022 paper that uses AI-generated critiques and revisions guided by a written set of principles (the "const...
-
A content filter is a rule-based or model-based system that blocks LLM outputs (or inputs) containing prohibited content categories — violence, sexual content, hate speech, self-harm, etc. — and is...
-
Continuous batching, sometimes called dynamic batching or in-flight batching, is an LLM serving technique where new incoming requests join a running batch immediately rather than waiting for the cu...
-
Data parallelism is the simplest and most common form of distributed training, where the model is replicated on each GPU and different micro-batches of training data are processed simultaneously ac...
-
A datasheet for a dataset is a structured documentation artifact for an ML training or evaluation dataset, proposed by Gebru et al. in a 2018 paper "Datasheets for Datasets" and modeled on the elec...
-
DeepSpeed is an open-source deep learning optimization library released by Microsoft Research in 2020 that provides memory-efficient training, distributed inference, and a suite of techniques for s...
-
DPO, short for Direct Preference Optimization, is an alignment technique introduced by Rafailov et al. in a May 2023 paper that achieves RLHF -quality results without training a separate reward mod...
-
A draft model is the small fast model used in speculative decoding to propose candidate tokens that the larger target model then verifies in parallel. The draft model must be from the same model fa...
-
Encoder-decoder architecture is the original Transformer configuration from the 2017 "Attention Is All You Need" paper, where an encoder processes the input sequence into a contextualized represent...
-
An evaluation card is a structured documentation artifact for an AI model evaluation, describing the benchmark used, the evaluation methodology, the prompts and conditions, the results, and the lim...
-
The feed-forward network, abbreviated FFN, is the second sublayer in each Transformer block (the first being multi-head attention ), responsible for applying nonlinear transformations to each posit...
-
FlashAttention is an exact attention algorithm introduced by Tri Dao et al. in a 2022 paper that dramatically accelerates self-attention by tiling computations to keep intermediate tensors in fast ...
-
FSDP, short for Fully Sharded Data Parallel, is a distributed training technique built into PyTorch since version 1.11 (2022) that shards model parameters, gradients, and optimizer states across mu...
-
Full fine-tuning is the training approach that updates all of a pretrained model's weights on task-specific data, in contrast to PEFT methods like LoRA , adapter layers, and prefix tuning that upda...
-
Function calling is the structured tool-use capability built into modern LLMs where the model emits a JSON object specifying a function name and arguments rather than free-form text, enabling relia...
-
GGUF, short for GPT-Generated Unified Format, is the binary file format used by Llama.cpp and Ollama to store quantized LLM weights, metadata, and tokenizer configuration in a single self-contained...
-
GPTQ is a one-shot quantization technique introduced by Frantar et al. in 2022 that produces 3-bit or 4-bit quantized LLMs with minimal accuracy loss using approximate second-order information (an ...
-
Gradient accumulation is a training technique that simulates larger effective batch sizes by accumulating gradients over multiple forward-backward passes before applying an optimizer step, allowing...
-
Gradient checkpointing is a memory-saving training technique that trades compute for memory by recomputing intermediate activations during the backward pass rather than storing them from the forwar...
-
Gradient descent is the iterative optimization algorithm that updates a neural network's weights by stepping in the direction opposite to the gradient of the loss function, gradually moving toward ...
-
Grouped-Query Attention, abbreviated GQA, is a multi-head attention variant introduced by Ainslie et al. in a 2023 Google paper that shares key and value projections across groups of query heads, r...
-
GSM8K, short for Grade School Math 8K, is a benchmark introduced by OpenAI in 2021 containing 8,500 grade-school-level multi-step math word problems requiring 2-8 steps of reasoning to solve. The b...
-
Guardrails are programmable policy enforcement layers around LLM applications that validate inputs and outputs against rule-based, regex, classifier, or LLM-judged criteria. Unlike model-level refu...
-
HELM, short for Holistic Evaluation of Language Models, is a benchmark framework introduced by Stanford's Center for Research on Foundation Models (CRFM) in 2022 that evaluates LLMs across 7 metric...
-
Helpful, Harmless, and Honest, often abbreviated HHH, is the canonical three-part value framework introduced by Anthropic in a 2021 paper ("A General Language Assistant as a Laboratory for Alignmen...
-
HumanEval is a code-generation benchmark introduced by OpenAI alongside Codex in a 2021 paper, containing 164 hand-written Python programming problems with unit tests for automatic verification. Ea...
-
Indirect prompt injection is a specific class of prompt injection attacks where the malicious instructions are embedded in third-party content the LLM processes — web pages it browses, documents in...
-
Instruction tuning is the post-pretraining adaptation technique that teaches a base LLM to follow natural-language instructions by training it on datasets of (instruction, response) pairs. The tech...
-
A jailbreak is an attack on an LLM that bypasses the model's safety training to elicit responses the model was trained to refuse — typically harmful instructions, restricted content, or operations ...
-
JSON mode is an LLM output constraint that forces the model to produce valid JSON, eliminating the parsing failures that plague free-text JSON generation. OpenAI introduced JSON mode in November 20...
-
KTO, short for Kahneman-Tversky Optimization, is an alignment technique introduced by ContextualAI in 2024 that draws on prospect theory from behavioral economics to align LLMs using unpaired good ...
-
Layer normalization, often abbreviated LayerNorm, is a normalization technique introduced by Ba, Kiros, and Hinton in 2016 that normalizes activations across the feature dimension within each token...
-
Learning rate is the hyperparameter that controls how large a step the optimizer takes in the direction of the negative gradient during training — too small and training is slow, too large and trai...
-
Llama.cpp is an open-source LLM inference engine written in pure C/C++ by Georgi Gerganov, released in March 2023, that enables CPU and consumer-GPU inference of quantized LLMs with minimal depende...
-
LoRA, short for Low-Rank Adaptation, is a parameter-efficient fine-tuning technique introduced by Microsoft Research in a 2021 paper by Hu et al. that has become the dominant approach to adapting l...
-
LoRA rank, often denoted r, is the dimensionality of the low-rank decomposition that LoRA adapters use to approximate weight updates — a key hyperparameter that controls the trade-off between adapt...
-
Megatron-LM is an open-source LLM training framework developed by NVIDIA, originally introduced in 2019 and expanded with the Megatron-Turing NLG 530B collaboration with Microsoft. The framework pr...
-
Mixed precision training is a technique that uses lower-precision floating point (typically FP16 or BF16) for most computations while keeping a master copy of weights in full FP32 precision, dramat...
-
MMLU, short for Massive Multitask Language Understanding, is a benchmark introduced by Hendrycks et al. in 2020 that tests LLM knowledge and reasoning across 57 subjects ranging from elementary mat...
-
A model card is a structured documentation artifact for an AI model covering its intended uses, training data, evaluation results, limitations, and ethical considerations, proposed by Mitchell et a...
-
MRKL, short for Modular Reasoning, Knowledge and Language (pronounced "miracle"), is an agent architecture introduced by AI21 Labs in 2022 that combines an LLM with external expert modules for task...
-
MT-Bench is a benchmark for evaluating chat-tuned LLMs on multi-turn conversational tasks, introduced by LMSYS in 2023 alongside the Chatbot Arena leaderboard. The benchmark contains 80 multi-turn ...
-
MTEB, short for Massive Text Embedding Benchmark, is the standard benchmark for evaluating embedding models , introduced by Hugging Face and Cohere researchers in 2023. The benchmark covers 8 task ...
-
Multi-head attention is the parallel-attention scheme in the Transformer that runs multiple self-attention operations in parallel with different linear projections of the input, then concatenates t...
-
Multi-Query Attention, abbreviated MQA, is an extreme variant of Grouped-Query Attention introduced by Shazeer in a 2019 paper, where all query heads share a single key and value projection. MQA re...
-
Ollama is an open-source LLM serving wrapper around Llama.cpp that adds a clean REST API, a Docker-Hub-style model registry, and a one-line install experience, making local LLM inference accessible...
-
ORPO, short for Odds Ratio Preference Optimization, is an alignment technique introduced by Hong et al. in early 2024 that merges SFT and preference optimization into a single training stage, elimi...
-
PagedAttention is the memory management algorithm at the heart of vLLM , introduced in a 2023 paper by UC Berkeley researchers that solved one of the most painful problems in LLM serving: KV cache ...
-
PEFT, short for Parameter-Efficient Fine-Tuning, is the umbrella term for a family of techniques that adapt large pretrained models by training only a tiny fraction of their parameters — typically ...
-
Perplexity is the exponential of the average negative log-likelihood that a language model assigns to a held-out text corpus, with lower perplexity indicating the model finds the text more probable...
-
Pipeline parallelism is a distributed training technique that partitions a neural network's layers across multiple GPUs or nodes, with each device handling a contiguous slice of the model. During t...
-
Plan-and-Execute is an agentic pattern where an LLM first generates a complete plan as a list of steps, then executes each step in sequence — separating the planning phase from execution rather tha...
-
Positional encoding is the mechanism that gives Transformer models information about the order of tokens in a sequence, since self-attention by itself is permutation-equivariant and cannot distingu...
-
Prefix caching is an LLM inference optimization that reuses the computed KV cache for shared prompt prefixes across multiple requests, eliminating redundant computation when many requests share the...
-
Prefix tuning is a PEFT technique introduced by Li and Liang (2021) that prepends a small sequence of learned continuous vectors (the prefix) to every layer's attention input, allowing task adaptat...
-
Prompt injection is a class of attacks where an attacker inserts instructions into LLM input that override or subvert the application's intended behavior, named by analogy to SQL injection. Direct ...
-
Prompt tuning, also called soft prompt tuning, is a PEFT technique introduced by Lester, Al-Rfou, and Constant (2021) that prepends a short sequence of learned continuous vectors directly to the in...
-
QLoRA, short for Quantized Low-Rank Adaptation, is an extension of LoRA introduced by Dettmers et al. in 2023 that combines 4-bit quantization of the base model with LoRA adapter training, enabling...
-
RAGAS, short for Retrieval-Augmented Generation Assessment, is an evaluation framework specifically designed for RAG pipelines, released as an open-source library in 2023 and adopted by many enterp...
-
ReAct, short for Reasoning and Acting, is an agentic LLM framework introduced by Yao et al. in a 2022 Google paper that interleaves reasoning traces with action calls, producing more reliable agent...
-
Red teaming for AI is the practice of having dedicated adversarial testers attempt to elicit harmful, biased, false, or otherwise problematic outputs from an LLM before deployment, modeled on red-t...
-
Reflexion is an agent pattern introduced by Shinn et al. in a 2023 paper that adds explicit self-reflection and learning from past attempts, enabling agents to improve performance on repeated trial...
-
Refusal training is the post-training technique that teaches LLMs to decline requests for harmful, dangerous, or policy-violating content — a fundamental component of every commercial LLM 's safety...
-
Residual connections, also called skip connections, are direct paths from a layer's input to its output that bypass the intermediate computation, introduced by He et al. in the 2015 ResNet paper an...
-
RLHF, short for Reinforcement Learning from Human Feedback, is the alignment technique introduced by OpenAI in InstructGPT (2022) and used to train ChatGPT, Claude, Gemini, and most major commercia...
-
RMSNorm, short for Root Mean Square Layer Normalization, is a simplified layer normalization variant introduced by Zhang and Sennrich in a 2019 paper that omits the mean-centering step and bias par...
-
RoPE, short for Rotary Position Embedding, is a positional encoding technique introduced by Su et al. in a 2021 paper that encodes absolute position via rotation matrices applied to query and key v...
-
ROUGE, short for Recall-Oriented Understudy for Gisting Evaluation, is a family of metrics for evaluating automatic summarization introduced by Chin-Yew Lin in 2004. The most-reported variants are ...
-
A safety classifier is a smaller specialized model that screens LLM inputs and outputs for harmful, toxic, or policy-violating content, typically deployed as a pre- or post-processing layer around ...
-
Self-Ask is an agentic prompting pattern introduced by Press et al. in a 2022 paper that improves multi-hop question-answering by having the LLM explicitly decompose complex questions into simpler ...
-
Self-attention is the core mechanism of the Transformer architecture, allowing each position in a sequence to attend to every other position when computing its representation. The mechanism compute...
-
SFT, short for Supervised Fine-Tuning, is the standard training phase that adapts a base LLM to follow instructions by training it on labeled examples of (input, desired output) pairs. SFT is typic...
-
Sparse attention is a family of self-attention variants that compute attention over only a structured subset of query-key pairs rather than the full quadratic set, dramatically reducing compute and...
-
Speculative decoding is an LLM inference acceleration technique introduced by Google researchers in 2022 and refined by DeepMind in 2023 that uses a small fast "draft model" to propose multiple can...
-
Structured output is the broader class of LLM output constraints that force responses to conform to a specified schema — typically JSON Schema, a Pydantic class, a Zod schema, or a regex pattern — ...
-
SwiGLU, short for Swish-Gated Linear Unit, is a feed-forward network variant introduced by Shazeer in a 2020 paper that combines the Swish activation function with a Gated Linear Unit (GLU) structu...
-
A system card is an extension of the model card concept that documents an end-to-end AI system — including the model, safety layers, deployment context, and known risks — rather than just the under...
-
Tensor parallelism is a distributed training technique that splits individual layer computations across multiple GPUs, typically within a single node where high-bandwidth interconnects like NVLink ...
-
TensorRT-LLM is NVIDIA's open-source LLM inference framework, released in late 2023, that compiles transformer models into highly optimized CUDA kernels for the lowest possible latency on NVIDIA GP...
-
Tool use protocol refers to the formal interface by which an LLM agent invokes external tools, including the schema for declaring available tools, the format for invoking them, and the convention f...
-
The Transformer is the neural network architecture introduced in the seminal 2017 paper "Attention Is All You Need" by Vaswani et al. at Google, replacing recurrent networks (LSTMs, GRUs) as the fo...
-
Triton Inference Server is NVIDIA's open-source model-serving framework, originally released in 2018, that serves any AI model — LLMs , vision, audio, classical ML — through a unified HTTP/gRPC API...
-
TruthfulQA is a benchmark introduced by Lin, Hilton, and Evans in 2021 that tests whether LLMs avoid generating false answers to questions designed to elicit common misconceptions, conspiracy theor...
-
vLLM is an open-source LLM inference engine released by UC Berkeley researchers in 2023 that has become the dominant high-throughput serving framework for self-hosted LLM deployments. The framework...
-
ZeRO, short for Zero Redundancy Optimizer, is a memory optimization technique introduced by Microsoft Research in 2019 that shards optimizer states, gradients, and (optionally) model parameters acr...
-
Angular distance is a variant of cosine similarity that converts the cosine value into a proper distance metric satisfying the triangle inequality, computed as arccos(cosine_similarity) divided by ...
-
Annoy, an acronym for Approximate Nearest Neighbors Oh Yeah, is an open-source ANN library released by Spotify in 2015 that uses random projection trees for efficient similarity search. The algorit...
-
Approximate Nearest Neighbor search, abbreviated ANN, is a class of algorithms that find vectors close to a query vector without guaranteeing they are the absolute closest, trading a small accuracy...
-
Batch embedding is the operation of generating embeddings for many input texts in a single API call or inference pass, dramatically improving throughput compared to processing one input at a time. ...
-
A bi-encoder, also called a dual-encoder, is the standard architecture for fast embedding -based retrieval: separate neural network instances encode the query and the documents independently into v...
-
Binary embeddings represent each vector dimension as a single bit (1 or 0) rather than a 32-bit float, achieving 32x storage compression and dramatically faster Hamming-distance retrieval at the co...
-
The BOS token (Beginning Of Sequence) is a special token that marks the start of model input, signaling to the model that what follows is the beginning of a new generation context rather than a con...
-
Byte-Pair Encoding, abbreviated BPE, is a tokenization algorithm originally developed as a data compression technique in 1994 and adapted to NLP by Sennrich, Haddow, and Birch in 2015. BPE starts w...
-
Chroma is an open-source vector database released in 2022 by Chroma Inc., designed specifically for the LLM application developer with a minimal API and zero-config local mode that runs entirely in...
-
Chunk overlap is the number of tokens or characters shared between consecutive chunks in a document, designed to prevent semantic information from being split across chunk boundaries and lost to re...
-
Chunk size is the target length of each document chunk, measured in tokens, characters, or sentences depending on the chunker. Production RAG systems typically use chunk sizes in the 200 to 1000 to...
-
Chunking is the process of splitting long documents into smaller pieces that can be individually embedded and retrieved by a RAG system, one of the most consequential design decisions in any retrie...
-
A code splitter is a structure-aware chunking tool that respects programming language syntax — function boundaries, class boundaries, comment blocks, import sections — when dividing source code int...
-
Cohere Rerank is a managed reranking service released by Cohere in 2023, providing pretrained cross-encoder rerankers that can rescore retrieval candidates with near-state-of-the-art accuracy via a...
-
Cold start indexing is the operation of building a vector index from scratch — ingesting documents, generating embeddings , and constructing the search structure — typically performed when a new RA...
-
Cosine similarity measures the angle between two vectors regardless of their magnitudes, computed as the dot product divided by the product of the vector norms. The result ranges from -1 (opposite ...
-
A cross-encoder reranker is a transformer model that takes a query and a candidate document together as joint input and outputs a single relevance score, in contrast to bi-encoder retrievers that s...
-
The curse of dimensionality refers to the collection of counter-intuitive phenomena that arise when dealing with very high-dimensional spaces, originally named by Richard Bellman in 1957 in the con...
-
Dimensionality reduction is the process of transforming high-dimensional data into a lower-dimensional representation while preserving as much relevant structure as possible. Classic linear methods...
-
DiskANN is a disk-resident graph-based ANN algorithm introduced in a 2019 Microsoft Research paper that enables billion-scale vector search on a single machine with modest RAM by storing most of th...
-
Document parsing is the process of extracting structured text and metadata from binary or formatted documents — PDFs, Word files, PowerPoint decks, HTML pages, emails — as the preprocessing step be...
-
Dot product similarity, also called inner product similarity, computes the sum of element-wise multiplications between two vectors, returning a scalar that grows with both the alignment and the mag...
-
Elasticsearch Vector refers to the dense vector field type and k-NN search capabilities added to Elasticsearch starting in version 7.3 and substantially enhanced in versions 8.x with HNSW indexing ...
-
An embedding API is a network endpoint that converts text or other modal input into embedding vectors via a remote model service, the most common pattern for accessing embedding models in productio...
-
Embedding dimension is the number of components in a vector produced by an embedding model , a fundamental architectural property that affects accuracy, storage cost, and retrieval performance. Com...
-
The EOS token (End Of Sequence) is a special token that signals the end of model output, indicating to the inference engine that generation should stop. When an LLM produces an EOS token during aut...
-
Euclidean distance, also called L2 distance, measures the straight-line distance between two vectors in n-dimensional space, computed as the square root of the sum of squared element-wise differenc...
-
Exact Nearest Neighbor search, sometimes called brute-force or flat search, computes the distance from the query to every vector in the collection and returns the truly closest matches — guaranteei...
-
FAISS, short for Facebook AI Similarity Search, is an open-source library released by Meta AI Research in 2017 that provides highly optimized C++ implementations of dozens of vector indexing and se...
-
Filtered vector search combines vector similarity search with structured predicates over metadata fields, returning the most similar vectors that also satisfy the filter conditions. Filters might b...
-
Float32 vectors represent each embedding dimension as a 32-bit IEEE 754 floating-point number, the default precision for most neural network outputs and the standard storage format in vector databa...
-
Hamming distance counts the number of positions at which two equal-length binary vectors differ, a natural metric for binary embeddings and hash codes. The distance is computed by XOR-ing the two v...
-
High-dimensional space refers to vector spaces with many independent axes — typically dozens, hundreds, or thousands — where geometric and statistical intuitions from two and three dimensions break...
-
HNSW, short for Hierarchical Navigable Small World, is a graph-based ANN algorithm introduced in a 2016 paper by Malkov and Yashunin that has become the dominant index type in modern vector databas...
-
Hybrid search is the combination of dense vector similarity search with sparse keyword search (typically BM25 or SPLADE ), merging the strengths of both retrieval modes for higher overall quality. ...
-
Index build time is the wall-clock duration required to construct a vector index from a collection of embeddings , an important operational metric that varies dramatically across ANN algorithms. HN...
-
Index refresh is the operation of incorporating newly ingested vectors into a queryable index, ranging from streaming near-real-time updates to scheduled batch rebuilds depending on the platform an...
-
Index sharding is the technique of partitioning a large vector index across multiple machines or storage nodes, with each shard holding a subset of the vectors and answering queries against its sub...
-
Inner product, also called dot product when applied to vectors, is the most fundamental similarity computation in machine learning — the sum of element-wise products of two vectors. Inner product i...
-
IVF, short for Inverted File Index, is an ANN algorithm that partitions the vector space into clusters (usually via k-means) at index time, then at query time searches only the clusters whose centr...
-
IVF-PQ combines the IVF clustering approach with Product Quantization compression, producing one of the most memory-efficient ANN index types available for large-scale vector search. IVF first part...
-
Jaccard similarity measures the overlap between two sets as the size of their intersection divided by the size of their union, producing a value between 0 (disjoint sets) and 1 (identical sets). Th...
-
LanceDB is an open-source serverless vector database released in 2023 by the team behind the Lance columnar format, designed specifically for multimodal AI workloads that combine text, image, audio...
-
Layout-aware parsing is the class of document parsing techniques that preserve and exploit visual structure — reading order, headings, tables, figures, columns, headers, footers — rather than treat...
-
LSH, short for Locality Sensitive Hashing, is one of the oldest families of ANN algorithms, dating back to a seminal 1998 paper by Indyk and Motwani that defined the formal framework. LSH uses hash...
-
Mahalanobis distance is a statistical distance metric that accounts for the covariance structure of a dataset — vectors are weighted such that variance-heavy dimensions contribute less than varianc...
-
Manhattan distance, also called L1 distance, taxicab distance, or city-block distance, measures the sum of absolute element-wise differences between two vectors — visualized as the path a taxi woul...
-
A markdown splitter is a structure-aware chunking tool that respects markdown syntax — headings, lists, code blocks, tables, and blockquotes — when dividing documents into chunks. Rather than split...
-
Marqo is an open-source end-to-end vector search engine that bundles embedding generation, vector indexing, and search into a single API, abstracting away the typical multi-component RAG architectu...
-
Matryoshka Representation Learning, often called Matryoshka embeddings, is a training technique introduced in a 2022 paper by Kusupati et al. that produces embedding vectors usable at multiple dime...
-
Milvus is an open-source vector database originally created by Zilliz in 2019 and now a graduated project of the LF AI & Data Foundation, designed for trillion-scale vector search across distribute...
-
NSG, short for Navigating Spreading-out Graph, is a graph-based ANN algorithm introduced in a 2018 paper by Fu et al. that achieves competitive recall and latency with HNSW while consuming substant...
-
OpenSearch k-NN is the vector search capability built into OpenSearch, the AWS-led open-source fork of Elasticsearch created in 2021. The k-NN plugin supports HNSW, IVF, and brute-force search impl...
-
The padding token (PAD) is a special token used to fill batches of inputs to the same length so that they can be processed together efficiently on parallel hardware like GPUs and TPUs. Without padd...
-
PCA, short for Principal Component Analysis, is a classical linear dimensionality reduction technique introduced by Karl Pearson in 1901 that projects high-dimensional data onto a lower-dimensional...
-
PDF extraction is the specific case of document parsing applied to PDF files, complicated by the fact that PDFs are a visual layout format rather than a semantic text format. Text in PDFs may be st...
-
pgvector is an open-source PostgreSQL extension first released in 2021 that adds native vector data types and similarity search to the world's most popular relational database. With pgvector instal...
-
Pinecone is a managed cloud-native vector database launched in 2021, designed for production-grade RAG , semantic search, and recommendation workloads with serverless scalability and sub-hundred-mi...
-
Product Quantization, abbreviated PQ, is a vector compression technique introduced by Jégou, Douze, and Schmid in 2011 that divides a high-dimensional vector into m subvectors and represents each s...
-
Qdrant is an open-source vector database written in Rust, first released in 2021 and known for high-performance HNSW indexing, rich payload filtering, and quantization options that dramatically red...
-
Recall@k is the standard evaluation metric for ANN algorithms, measuring the fraction of true top-k nearest neighbors (as computed by exact search) that the approximate algorithm actually returns. ...
-
Reciprocal Rank Fusion, abbreviated RRF, is a simple but remarkably effective algorithm for combining ranked result lists from multiple retrieval systems into a single unified ranking. RRF assigns ...
-
Recursive chunking is a hierarchical splitting strategy that attempts to split text at meaningful natural boundaries first, falling back to coarser boundaries only when chunks exceed the target siz...
-
Redis Vector Search refers to the vector similarity search capabilities added to Redis Stack and Redis Enterprise starting in 2022, leveraging the FT.SEARCH command of the RediSearch module to perf...
-
A reranker is a second-stage retrieval component that takes an initial candidate set (typically the top-k results from a faster first-stage retriever) and produces a refined ranking using a more ac...
-
Scalar Quantization, abbreviated SQ, is a vector compression technique that maps each float32 component of a vector to a lower-precision representation — typically int8 (8 bits) or even binary (1 b...
-
ScaNN, short for Scalable Nearest Neighbors, is an open-source ANN library released by Google Research in 2020 that combines anisotropic vector quantization with optimized SIMD search to deliver st...
-
Semantic chunking is a content-aware splitting strategy that uses embeddings to detect topic shifts within a document and split at those natural semantic boundaries, rather than at fixed token coun...
-
Sentence splitting is the preprocessing step that segments text into individual sentences as a foundation for downstream chunking , embedding , and retrieval workflows. Naive sentence splitting on ...
-
SentencePiece is an open-source subword tokenization library developed at Google and released in 2018, designed to be language-agnostic by treating input as a raw byte stream rather than relying on...
-
Sliding window is a chunking strategy that produces a sequence of overlapping chunks by sliding a fixed-size window across the document with a fixed stride smaller than the window size. For example...
-
Special tokens are reserved entries in a tokenizer vocabulary that carry structural or semantic meaning rather than representing ordinary text content. Common special tokens include the beginning-o...
-
Squared Euclidean distance is the sum of squared element-wise differences between two vectors, equivalent to Euclidean distance without the final square root operation. Skipping the square root sav...
-
Subword tokenization is the approach of splitting text into units smaller than whole words but larger than individual characters, the dominant tokenization strategy in modern LLMs . Common subword ...
-
tiktoken is an open-source Python library released by OpenAI in 2022 that implements the BPE tokenizers used by GPT-2, GPT-3, GPT-3.5, GPT-4, GPT-4o, and the o-series reasoning models. The library ...
-
A token is the basic unit of input and output for a language model — typically a subword fragment, occasionally a whole word, sometimes a single character, depending on the tokenizer. In modern LLM...
-
A token ID is the integer index of a token within its tokenizer vocabulary , the actual numerical representation that LLMs process internally. When text enters a model it is first tokenized into a ...
-
Tokenization is the process of splitting raw text into discrete units called tokens that a language model can process, the foundational preprocessing step for every text-based AI system. Tokens are...
-
A tokenizer vocabulary is the fixed set of token strings (or byte sequences) and their integer IDs that a tokenizer can produce, typically containing 30,000 to 200,000 entries depending on the mode...
-
Top-k retrieval is the parameter that controls how many of the most similar vectors a vector search query returns, with k typically ranging from 1 to a few hundred depending on the workload. Smalle...
-
t-SNE, short for t-Distributed Stochastic Neighbor Embedding, is a nonlinear dimensionality reduction algorithm introduced by van der Maaten and Hinton in 2008 that has become the standard techniqu...
-
UMAP, short for Uniform Manifold Approximation and Projection, is a nonlinear dimensionality reduction algorithm introduced by McInnes, Healy, and Melville in 2018 that has largely replaced t-SNE f...
-
Unigram tokenization, more precisely the Unigram Language Model tokenization algorithm, is a probabilistic alternative to BPE introduced by Taku Kudo in 2018 and made available through SentencePiec...
-
The UNK token (UNKnown) is a special token that older tokenizers used to represent input characters or sequences that fell outside the vocabulary. UNK was prevalent in pre-2018 NLP systems where to...
-
Upsert is the combined insert-or-update operation that adds new vectors to a vector collection or replaces existing vectors with the same ID. The term, common across most modern vector databases , ...
-
Vald is an open-source distributed vector search engine released by Yahoo Japan in 2020, designed for cloud-native deployment on Kubernetes with horizontal scaling, automatic indexing, and self-hea...
-
Vector cache is a memory-resident store of frequently used embedding vectors, query results, or intermediate computations, designed to reduce latency and cost in production RAG pipelines. Caching c...
-
A vector collection (sometimes called an index, class, or table depending on the platform) is the logical container inside a vector database that holds a set of related embeddings along with their ...
-
A vector database is a specialized data store designed to index, search, and manage high-dimensional numerical vectors — the embeddings produced by neural networks — at scale. Unlike traditional re...
-
Vector dimensionality refers to the number of independent axes in a vector space, a concept that drives both expressive power and computational complexity in embedding -based retrieval. Higher-dime...
-
A vector filter is a structured predicate combined with a vector similarity query that restricts results to vectors matching specific metadata conditions — for example, retrieving the most similar ...
-
A vector index is the in-memory or on-disk data structure inside a vector database that organizes embeddings for fast approximate nearest neighbor search. Without an index, finding the closest vect...
-
A vector namespace is a logical partition within a vector collection that isolates a subset of vectors for multi-tenant or multi-context scenarios, allowing queries to be scoped to a single tenant'...
-
Vector normalization is the operation of rescaling a vector to unit length (L2 norm = 1) by dividing each component by the original Euclidean magnitude. Normalized vectors live on the surface of th...
-
Vector replication is the practice of maintaining multiple copies of a vector index across machines, regions, or data centers for high availability, disaster recovery, and read scaling. Replicated ...
-
A vector schema is the formal definition of the vector dimension, distance metric, metadata fields, and index parameters that govern how vectors are stored, validated, and queried in a vector colle...
-
A vector search query is a request to a vector database that supplies a query vector and asks for the top-k closest stored vectors by some similarity metric, optionally with structured filters. Vec...
-
Vespa is an open-source big data serving engine originally developed at Yahoo and open-sourced in 2017, combining vector search, structured retrieval, full-text search, and inference of machine lea...
-
Weaviate is an open-source vector database first released in 2019 by the company SeMI Technologies (now Weaviate B.V.), notable for combining vector search with a built-in GraphQL API and a strong ...
-
WordPiece is a subword tokenization algorithm introduced by Google in 2012 and popularized by BERT in 2018, similar to BPE but with a different merge criterion based on maximum likelihood rather th...
-
Abstractive Summarization generates new sentences that capture the meaning of source material — paraphrasing, restructuring, and synthesizing rather than copying existing text. The approach mirrors...
-
The Action Space is the set of all possible actions an AI agent can take in its environment — every tool it can call, every API it can hit, every command it can issue. The concept comes from reinfo...
-
AGI (Artificial General Intelligence) refers to AI systems that match or exceed human-level cognitive ability across the full range of cognitive tasks — including reasoning, learning, planning, cre...
-
An AI Skill is a packaged, reusable capability that combines a prompt, a knowledge base, tools, and configuration into a single discoverable unit — like a function in software, but for AI capabilit...
-
ASI (Artificial Superintelligence) refers to hypothetical AI systems that dramatically exceed human-level intelligence across all cognitive domains — not merely matching humans but surpassing them ...
-
Auto-Categorization assigns category labels from a predefined hierarchy or taxonomy to content automatically — using AI to scale classification beyond what humans can review manually. Common scenar...
-
Auto-Labeling uses AI itself to generate training labels — bootstrapping supervised learning at scales human annotation cannot match. Common techniques include weak supervision (Snorkel and similar...
-
Auto-Tagging applies metadata labels to content automatically using AI — classifying documents, images, emails, support tickets, and database records at scale. Common auto-tagging targets include t...
-
Batch Inference processes large groups of inputs together rather than one at a time — exchanging latency for throughput and efficiency. Common batch workloads include nightly scoring of every custo...
-
Beam Search is a classical decoding algorithm that explores multiple candidate output sequences in parallel — keeping the top K most likely partial sequences (the "beam") at each step and continuin...
-
BFloat16 (BF16, Brain Floating Point 16) is a 16-bit floating-point format that trades precision (fewer mantissa bits) for wider dynamic range (matching FP32's exponent range) — making it especiall...
-
BGE (BAAI General Embedding) is a family of open-source embedding models from the Beijing Academy of Artificial Intelligence — released under MIT license and broadly considered among the best open-...
-
BGE-M3 is BAAI's multi-functional embedding model — supporting three retrieval modes (dense, sparse, multi-vector) in a single model, three languages categories (100+ languages), and three input le...
-
A Browser Agent is an AI system that autonomously navigates the web — opening pages, clicking links, filling forms, extracting information, and completing multi-step tasks across websites. Browser ...
-
An AI Capability is a high-level competence that an AI system possesses — like "translate between languages," "summarize documents," "answer questions about products," "generate code from descripti...
-
A Capability Map is a structured visualization of an organization's AI capabilities — what the organization can do today, what it plans to build, and where strategic gaps exist. Capability maps are...
-
Claude 3 Opus is Anthropic's flagship model from the Claude 3 family, released in March 2024 with the strongest reasoning, instruction-following, and coding performance in the Claude 3 lineup. The ...
-
Claude 3.5 Haiku is Anthropic's small, fast model from November 2024 — designed for high-volume applications where speed and cost matter as much as quality. Priced at roughly $1 per million input t...
-
Claude 3.5 Sonnet is Anthropic's mid-tier model from June 2024 that became one of the most popular production LLMs for its combination of capability, speed, and cost. The model surpassed Claude 3 O...
-
Claude 4 Opus is Anthropic's flagship model in the Claude 4 family, succeeding Claude 3 Opus as the premium tier with substantially improved reasoning, coding, and agentic capabilities. The model i...
-
Claude 4 Sonnet is Anthropic's mid-tier model in the Claude 4 family — succeeding Claude 3.5 Sonnet as the production workhorse. The model balances capability and cost: bringing significant gains o...
-
Claude 4.5 Sonnet is Anthropic's enhanced mid-tier model, advancing the Claude 4 Sonnet line with notable improvements on coding, agentic, and tool-use benchmarks. The release positioned Claude 4.5...
-
Claude Opus 4.1 is an enhanced revision in Anthropic's Claude 4 family at the premium Opus tier — continuing the lineage of frontier capability with iterative improvements over Claude 4 Opus. The ....
-
CLIP (Contrastive Language-Image Pre-training) is OpenAI's foundational multimodal model released in 2021 — trained to produce joint embeddings of text and images in the same vector space, enabling...
-
Cloud Inference runs AI models on remote, scalable infrastructure provided by hyperscalers (AWS, Azure, Google Cloud), AI labs (OpenAI, Anthropic, Google AI, Cohere), or specialty providers (Togeth...
-
Code Completion suggests the next characters, lines, or blocks of code as developers type — the most common form of AI coding assistance and the foundation of products like GitHub Copilot, Tabnine,...
-
Code Generation produces software source code from natural-language descriptions, function signatures, comments, or example test cases. The capability has been transformed by LLMs trained extensive...
-
Cohere Embed v3 is Cohere's third-generation embedding model — designed specifically for enterprise retrieval applications with strong performance on production RAG workloads. The family includes v...
-
ColBERT (Contextualized Late Interaction over BERT) is a retrieval architecture introduced by Khattab and Zaharia in 2020 — pioneering a hybrid approach between bi-encoder embeddings (fast, less ac...
-
Command R+ is Cohere's enterprise-focused LLM released in April 2024 — designed specifically for retrieval-augmented generation (RAG), tool use, and multilingual enterprise workloads. The 104B-para...
-
Computer Use is an AI capability that lets a model interact with computer interfaces like a human user — moving cursors, clicking buttons, typing text, taking screenshots, navigating between applic...
-
Concept Extraction identifies abstract concepts mentioned in text — not just named entities but ideas, topics, themes, and domain-specific terminology. Where NER pulls out specific people, places, ...
-
Content Classification assigns categories from a predefined taxonomy to documents, emails, images, audio, or video — at scale and automatically. Common business uses include routing support tickets...
-
Context Stuffing is the practice of placing large amounts of information directly into an LLM's prompt — entire documents, long conversation histories, full knowledge bases — relying on the model's...
-
Coreference Resolution identifies when different expressions in text refer to the same entity — "Apple announced new products this week. The company unveiled iPhone 17 and Vision Pro 2. Tim Cook sa...
-
Cross-Modal Retrieval is the task of retrieving items from one modality using queries in a different modality — finding images that match a text description, finding audio clips that match a video ...
-
Data Annotation is the process of attaching labels, classifications, or metadata to raw data — creating the training and evaluation sets that supervised AI models depend on. Common annotation work ...
-
A Data Catalog is a centralized inventory of an organization's data assets — datasets, tables, dashboards, models, APIs — enriched with metadata, lineage, ownership, business context, and access co...
-
DBRX is Databricks' March 2024 release of a 132B-parameter mixture-of-experts open-weight model — positioned as a high-performance enterprise-friendly alternative to proprietary frontier models. Th...
-
DeepSeek R1 is DeepSeek's January 2025 reasoning-focused model that delivered o1-class reasoning performance under an open-weight license — making frontier reasoning capability available for self-h...
-
DeepSeek V3 is the Chinese AI lab DeepSeek's December 2024 release — a 671B-parameter mixture-of-experts model with 37B active parameters per token that demonstrated frontier-class capability at re...
-
A Dense Model is a neural network where all parameters participate in every forward pass — contrasting with sparse models (MoE, pruned networks) where only a fraction of parameters are active per i...
-
Document Summarization condenses long documents into shorter forms — abstracts, executive summaries, bullet-point recaps, or one-line headlines — while preserving the most important information. Th...
-
E5 (Embeddings from Bidirectional Encoder Representations) is Microsoft Research's family of open-source embedding models released throughout 2022-2024. The family includes E5-base, E5-large, E5-mi...
-
Edge Inference runs AI models directly on user devices — phones, laptops, browsers, IoT sensors, vehicles — instead of sending data to remote cloud servers. The approach reduces latency (no network...
-
Entity Extraction is the broader discipline of pulling structured information out of unstructured text — encompassing NER, attribute extraction, key-value extraction from forms, table extraction fr...
-
Entity Linking resolves mentions in text to specific records in a knowledge base — distinguishing whether "Apple" refers to the technology company, the fruit, or the record label. Entity linking sy...
-
Extractive Summarization selects the most important existing sentences from source material to form a summary — without generating new text. The approach guarantees faithfulness (every word came fr...
-
FP16 (Half-Precision Floating Point, also called IEEE float16) represents numbers using 16 bits — halving the memory and bandwidth requirements of standard 32-bit FP32 while still providing high ac...
-
A Function Schema describes the signature of a callable function or tool that an AI model can invoke — its name, purpose, input parameters with types and constraints, and expected output. Function ...
-
Gemini 1.5 Flash is Google's fast, lightweight variant in the Gemini 1.5 family — designed for high-volume, low-latency applications while retaining most of the long-context capability of Gemini 1....
-
Gemini 1.5 Pro is Google DeepMind's mid-2024 flagship model in the Gemini 1.5 family, notable for its breakthrough context window — initially 1 million tokens, later extended to 2 million tokens fo...
-
Gemini 2.0 Flash is Google DeepMind's late-2024 release marking the start of the Gemini 2 family — emphasizing speed, multimodal output (native image and audio generation), and improved tool-use ca...
-
Gemini 2.5 Pro is Google DeepMind's flagship in the Gemini 2.5 family, advancing the Pro tier with improved reasoning, coding, and agentic capabilities while maintaining the massive context window ...
-
GGUF (GPT-Generated Unified Format) is a file format developed by the llama.cpp project to store quantized large language models efficiently — replacing the older GGML format with better metadata, ...
-
GPT-4 is OpenAI's flagship large language model released in March 2023 — the model that brought multimodal frontier AI to mainstream enterprise use. The original GPT-4 introduced a step-change impr...
-
GPT-4 Turbo is OpenAI's enhanced variant of GPT-4 released in November 2023 at OpenAI's DevDay, bringing several improvements over the original GPT-4: a 128K context window (up from 32K), improved ...
-
GPT-4o ("o" for "omni") is OpenAI's natively multimodal model released in May 2024, accepting text, audio, image, and video inputs and producing text, audio, and image outputs from a single model. ...
-
GPT-4o mini is OpenAI's small, fast, low-cost variant of GPT-4o released in July 2024 — designed as a high-volume workhorse for tasks that don't need flagship intelligence. Priced at roughly $0.15 ...
-
GPT-5 is OpenAI's next-generation flagship model in the post-GPT-4 family, broadly anticipated through 2024-2025 with various previews and incremental releases. The model represents OpenAI's contin...
-
Greedy Decoding is the simplest LLM output strategy — always pick the single most likely next token. The approach is deterministic, fast, and reproducible, but tends to produce repetitive, bland ou...
-
Grok-3 is xAI's third-generation LLM, succeeding Grok-1 (open-sourced under Apache 2.0 in March 2024) and Grok-2 in advancing xAI's frontier-model capability. The model is featured prominently in X...
-
GTE (General Text Embeddings) is Alibaba's family of open-source embedding models — released under MIT license and ranking among the top open-source options on MTEB benchmark. The family includes G...
-
A Hard Prompt is a prompt expressed in actual natural-language tokens that a human can read, write, and edit — the default form of prompt in most AI applications. Hard prompts are interpretable, ea...
-
Hugging Face is the central platform of the open-source AI ecosystem — hosting more than a million AI models, datasets, and Spaces (live demos). The company started in 2016 as a chatbot app, pivote...
-
Image Captioning generates natural-language descriptions of images — converting visual content into text that humans and downstream AI systems can understand. The task is foundational to accessibil...
-
Inference is the runtime phase of an AI system — the stage where a trained model produces predictions, classifications, or generated content from new inputs. Where training is a one-time (or period...
-
Inference Acceleration encompasses the hardware, software, and model optimizations that make AI inference faster, cheaper, and more energy-efficient. Hardware accelerators include NVIDIA H100/H200 ...
-
An Inference API is a network-accessible service that runs AI models on behalf of clients — exposing trained models through REST or streaming endpoints so applications can use AI without managing m...
-
Inference Cost is the operational expense of running an AI system in production — typically charged per token for hosted LLMs or measured in compute hours for self-hosted models. Hosted-model prici...
-
An Inference Engine is the runtime software that loads a trained AI model and serves predictions to applications. Modern inference engines optimize for different priorities — throughput (batch jobs...
-
Inference Latency is the time between sending a prompt to an AI system and receiving its full response. Latency matters enormously for user experience — a chatbot that takes 30 seconds to respond f...
-
An Inference Pipeline is the end-to-end sequence of operations that transforms a user input into a final AI output — including pre-processing, retrieval, model invocation, post-processing, and resp...
-
INSTRUCTOR is an embedding-model family from researchers at Hong Kong University of Science and Technology — notable for accepting natural-language task instructions alongside text input. Instead o...
-
INT4 Quantization represents AI model weights using only 4 bits per parameter, compressing memory footprint by 8x compared to standard 32-bit floating point. A 70-billion-parameter model that requi...
-
INT8 Quantization represents AI model weights using 8-bit integers, halving memory compared to FP16 and quartering compared to FP32 — while maintaining nearly identical accuracy to the original mod...
-
Jina Embeddings is a family of open-source embedding models from Jina AI — notable for supporting very long input contexts and for the company's broader open-source AI ecosystem (Jina, Finetuner, D...
-
JSON Mode is an LLM feature that constrains output to valid JSON — preventing the malformed responses that plague applications relying on free-text generation. OpenAI introduced JSON mode in Novemb...
-
Keyword Extraction identifies the most important terms or phrases in a document — supporting search indexing, content tagging, summarization, SEO, and analytics. Classical approaches include TF-IDF...
-
Knowledge Distillation trains a smaller "student" model to mimic the behavior of a larger "teacher" model — transferring capability while dramatically reducing inference cost. The technique was int...
-
KV Cache (Key-Value Cache) stores the intermediate attention computations from previously-processed tokens so that LLM inference does not recompute them at every generation step. Without KV caching...
-
Latent Dirichlet Allocation (LDA) is the classic probabilistic algorithm for topic modeling, introduced by Blei, Ng, and Jordan in 2003. LDA represents each document as a mixture of topics and each...
-
Llama 3 is Meta's open-weight large language model family released in April 2024 — making serious frontier-quality models available for free download, modification, and self-hosting. Llama 3 launch...
-
Llama 3.1 is Meta's mid-2024 expansion of the Llama 3 family — bringing the 405B-parameter variant alongside refreshed 70B and 8B versions, plus longer 128K context windows across all sizes. The 40...
-
Llama 3.2 is Meta's September 2024 release adding multimodal vision capabilities to the Llama family for the first time. The release included 11B and 90B vision-language variants alongside smaller ...
-
Llama 3.3 is Meta's December 2024 release that delivered Llama 3.1 405B-level performance in the much smaller 70B parameter size — dramatically reducing the cost of running near-frontier models. Th...
-
Llama 4 is Meta's next major release in the Llama family, advancing the state of open-weight AI with improvements across reasoning, multimodality, context length, and agentic capabilities. The rele...
-
Long Context refers to LLMs that process very large input prompts — typically 100K tokens or more, up to multi-million-token models. Context windows have expanded dramatically: from 2K-4K in origin...
-
Lost in the Middle describes a documented failure mode of long-context LLMs: information placed in the middle of a long context window is sometimes ignored, while information at the start and end g...
-
Machine Translation (MT) converts text from one language to another automatically — a foundational AI task with decades of history. The field evolved from rule-based MT (1950s-80s) through statisti...
-
Metadata is structured information that describes data — author, creation date, source, format, classification, sensitivity, tags, relationships, lineage — without containing the data's primary con...
-
Metadata Enrichment is the process of automatically generating additional descriptive information about content — using AI to extract entities, classify topics, generate summaries, identify sentime...
-
A Meta-Prompt is a prompt that produces or improves other prompts — using an LLM to design, refine, or critique the prompts that will be used in production. The technique recognizes that LLMs thems...
-
MiniLM is Microsoft Research's family of small, fast embedding models that became foundational to the open-source embedding ecosystem. The all-MiniLM-L6-v2 variant — a 6-layer, 22M-parameter model ...
-
Mistral Large is the flagship model from French AI lab Mistral AI, positioned as a premium European-developed alternative to American frontier models. Mistral Large 2, released in July 2024, has 12...
-
Mistral Medium is Mistral AI's mid-tier model — balancing capability and cost for production workloads that don't require the flagship Mistral Large. The model serves as Mistral's everyday workhors...
-
Mixed Precision is a training and inference technique that uses different numerical precisions for different parts of a neural network — typically keeping critical operations (loss computation, gra...
-
Mixtral 8x22B is Mistral AI's April 2024 expansion of the Mixtral mixture-of-experts line — featuring 8 experts of 22B parameters each, activating 2 experts per token for inference equivalent to a ...
-
Mixtral 8x7B is Mistral AI's December 2023 release that brought sparse mixture-of-experts (MoE) architecture to the open-weight model ecosystem. The model has 8 expert sub-networks of 7B parameters...
-
Mixture of Experts (MoE) is a neural network architecture where many specialized "expert" sub-networks share the work of processing input — with a routing layer determining which experts handle eac...
-
Model Compression encompasses techniques that reduce an AI model's size and inference cost — including quantization, pruning, knowledge distillation, low-rank factorization, and neural architecture...
-
A Model Hub is a centralized repository for AI models — hosting weights, model cards, documentation, evaluation metrics, and usage instructions for many models in a searchable format. Hugging Face ...
-
Model Serving is the infrastructure layer that hosts trained AI models and exposes them to applications via APIs, streaming endpoints, or batch interfaces. Common serving frameworks include NVIDIA ...
-
A Multimodal Embedding is a vector representation that captures information from multiple modalities — text and images, text and audio, video and text — in a single shared space. Multimodal embeddi...
-
Named Entity Disambiguation (NED) resolves ambiguous entity mentions to specific real-world entities — distinguishing whether "Apple" refers to the technology company, the fruit, or the record labe...
-
Named Entity Recognition (NER) identifies and classifies named entities in text — people, organizations, locations, dates, monetary amounts, products, medical terms, legal citations — and remains o...
-
Narrow AI (also called Weak AI or Specialized AI) refers to AI systems designed to perform specific, well-defined tasks — image classification, language translation, game playing, fraud detection, ...
-
Needle in a Haystack is the standard benchmark for evaluating how well long-context LLMs retrieve specific information buried in a large context window. The benchmark places a specific fact (the "n...
-
Nemotron is NVIDIA's family of open-weight LLMs designed to demonstrate the capabilities of NVIDIA's training infrastructure and to provide enterprise-grade models for the NVIDIA AI ecosystem. Majo...
-
Neural Machine Translation (NMT) uses neural networks to translate between languages — the current dominant approach to machine translation, replacing earlier rule-based and statistical methods. NM...
-
Neural-Symbolic AI (NeSy) combines the strengths of neural networks (learning from data, handling perception, scaling to large datasets) with the strengths of symbolic AI (explicit reasoning, inter...
-
Nomic Embed is Nomic AI's family of open-source embedding models — notable for being among the first fully reproducible open-source embedders, with training code, training data, and model weights a...
-
o1 is OpenAI's first reasoning-focused model, introduced in September 2024 as a new model family distinct from the GPT line. Where GPT models respond quickly with relatively shallow reasoning, o1 u...
-
o1-mini is OpenAI's smaller, faster, cheaper reasoning model in the o-series family — released alongside o1 in September 2024. Where o1 targets the hardest reasoning tasks, o1-mini focuses on codin...
-
o3 is OpenAI's next-generation reasoning model, announced in December 2024 as the successor to o1 with dramatic performance gains on the hardest reasoning benchmarks. The model achieved a breakthro...
-
o3-mini is OpenAI's smaller variant of o3 — bringing extended reasoning capabilities at lower cost and latency. The model balances o3's reasoning approach with o1-mini's economic pricing, targeting...
-
o4-mini is the next iteration in OpenAI's small-reasoning-model line, advancing the cost-effective reasoning model that paired with each new flagship. Following the pattern of o1-mini and o3-mini, ...
-
OCR (Optical Character Recognition) converts text inside images — scanned documents, photographs of receipts, screenshots, photos of whiteboards — into machine-readable text. OCR has been a product...
-
ONNX Runtime is a high-performance cross-platform inference engine for models in the Open Neural Network Exchange (ONNX) format — a vendor-neutral standard for representing trained neural networks....
-
On-Premise Inference runs AI models entirely inside an organization's own datacenters or private cloud — keeping data, prompts, and outputs behind the corporate firewall. The approach is essential ...
-
An Ontology is a formal model of concepts in a domain and the relationships between them — more expressive than a taxonomy because it captures not just hierarchy but rich relationships ("is-a," "pa...
-
An Output Schema is a structured definition of the expected format of an LLM response — typically expressed in JSON Schema, Pydantic models, or framework-specific schema notation. Schemas constrain...
-
Pair Programming AI extends code completion into a conversational coding partner — explaining code, debugging issues, suggesting refactoring, writing tests, and engaging in technical discussion alo...
-
Persona Prompting establishes detailed character profiles for AI assistants — not just role but personality traits, communication style, areas of expertise, knowledge limits, ethical commitments, a...
-
Phi-3 is Microsoft's family of small language models (SLMs) released in April 2024 — emphasizing that careful data curation can produce capable models at parameter counts dramatically smaller than ...
-
Phi-4 is Microsoft's December 2024 successor to Phi-3 — a 14B-parameter small language model that demonstrated frontier-class reasoning capability through aggressive use of synthetic training data....
-
A Plugin Architecture allows third-party code or capabilities to extend a core platform — without modifying the core itself. In AI, plugin architectures let LLMs invoke external tools, APIs, data s...
-
Prefix Tuning is a parameter-efficient adaptation technique that prepends learned vectors to every layer of a transformer's attention mechanism — going deeper than soft prompts (which only affect t...
-
Prompt Caching stores the intermediate KV-cache state for frequently-reused prompt prefixes — dramatically reducing latency and cost when the same large prompt appears across many requests. The tec...
-
A Prompt Chain links multiple prompts into a sequence where each step's output feeds the next — breaking complex tasks into reasoning steps that smaller, focused prompts can each handle reliably. A...
-
Prompt Compression reduces the token count of a prompt while preserving its essential information — directly cutting cost and improving latency, often dramatically. Long prompts (retrieved document...
-
Prompt Decomposition breaks a complex user request into smaller sub-questions that the AI can answer more reliably — then combines the answers into a final response. The technique is especially pow...
-
A Prompt Library is an organized, governed collection of prompt templates available across an enterprise — eliminating duplicate work, propagating best practices, and ensuring consistency across te...
-
A Prompt Manager is the operational system that stores, versions, deploys, and monitors prompts across an enterprise — providing the runtime layer for prompts the way databases provide the runtime ...
-
Prompt Optimization is the systematic process of improving prompts to maximize output quality, minimize cost, or balance both — using techniques ranging from manual iteration to automated search. C...
-
A Prompt Template is a reusable, parameterized prompt structure with placeholders that get filled in at runtime — separating prompt design from prompt usage. Templates make AI applications maintain...
-
Prompt Tuning is a parameter-efficient adaptation technique that learns a small number of "soft prompt" embeddings to steer a frozen large language model toward a specific task — without modifying ...
-
Prompt Versioning treats prompts like software artifacts — tracking every change, attaching metadata about who changed what when and why, and enabling rollback when a new version performs worse tha...
-
Pruning removes unnecessary connections or parameters from a neural network, shrinking model size and accelerating inference while preserving most accuracy. Common approaches include magnitude prun...
-
P-Tuning is a parameter-efficient fine-tuning technique that uses a small trainable neural network (a "prompt encoder") to generate the soft-prompt embeddings — making prompt tuning more expressive...
-
Quantization reduces the precision of an AI model's numerical weights — typically from 32-bit floating point (FP32) down to 16-bit (FP16), 8-bit integer (INT8), 4-bit (INT4), or even lower — dramat...
-
Qwen 2.5 is Alibaba's open-weight LLM family released in September 2024 — including variants from 0.5B to 72B parameters plus specialized versions for coding (Qwen2.5-Coder) and math (Qwen2.5-Math)...
-
Qwen 3 is Alibaba's next-generation open-weight model family, advancing the Qwen line with improved reasoning, coding, agentic, and multimodal capabilities. The release continues Alibaba's strategy...
-
Rate Limiting controls how many requests a client can make to an API in a given time window — protecting infrastructure from abuse, ensuring fair access among customers, and enforcing pricing tiers...
-
Real-Time Inference produces AI predictions within human-perceptible latency budgets — typically under one second for chat, under 100ms for code completion, and under 50ms for fraud detection or ad...
-
A Reasoning Model is an LLM specifically designed to perform extended internal reasoning before producing an output — exploring multiple paths, checking work, backtracking from errors, and synthesi...
-
Reflection Prompting asks the AI to evaluate its own response — checking for errors, missing information, logical inconsistencies, or quality issues — before delivering a final answer. The techniqu...
-
Relation Extraction identifies relationships between entities in text — "Apple acquired Beats Electronics in 2014" yields the relation (Apple, acquired, Beats Electronics, 2014). The technique is f...
-
Robotic Process Automation (RPA) automates repetitive business processes using software "bots" that interact with applications through their user interfaces — clicking buttons, copying data between...
-
Role Prompting instructs the AI model to adopt a specific persona or role — "You are an expert financial analyst," "You are a senior security engineer," "You are a kindergarten teacher" — to shape ...
-
A Schema is a formal description of data structure — the columns of a database table, the fields of a JSON object, the elements of an XML document, or the structure of a knowledge representation. S...
-
Self-Consistency is a reasoning technique that improves answer quality by sampling multiple chain-of-thought reasoning paths for the same question, then taking the majority answer. Introduced by Wa...
-
Self-Refine is an iterative improvement technique where an LLM generates an initial answer, critiques its own response, then revises it — repeating until the output meets quality criteria. Introduc...
-
Semantic Tagging applies meaningful, machine-readable labels to content based on understanding what the content actually means — going beyond surface keyword matching to capture concepts, themes, a...
-
Sentence-BERT (SBERT) is a foundational framework for producing sentence-level embeddings using BERT-based architectures — introduced by Reimers and Gurevych in 2019. The breakthrough was using Sia...
-
Sentiment Analysis classifies text by emotional valence — positive, negative, neutral, or finer-grained emotions like joy, anger, frustration, or satisfaction. Real-world applications include monit...
-
SigLIP (Sigmoid Loss for Language-Image Pre-training) is Google Research's 2023 CLIP-successor model that improved on CLIP's training methodology — using a simple sigmoid loss instead of softmax co...
-
Skill Authoring is the process of designing, writing, testing, and documenting an AI skill — the activity that produces new capabilities for users and applications to consume. Authoring involves mu...
-
A Skill Chain is a sequence of AI skills executed in order, with each step's output feeding the next — a specialized form of skill composition focused on linear workflows. Skill chains are the defa...
-
Skill Composition combines multiple AI skills into a workflow that accomplishes a larger task — chaining outputs to inputs, branching based on intermediate results, looping for iterative refinement...
-
Skill Deployment moves a skill from development into production — making it available to users and applications at scale. The process typically includes promotion through environments (dev → test →...
-
Skill Discovery is the process by which users, applications, or other AI agents find the right skill for a given task — through search, recommendations, semantic matching, or learned routing. As sk...
-
Skill Evaluation measures how well an AI skill performs against quality, safety, cost, and latency criteria — using automated metrics, human judgment, or LLM-as-judge approaches. Common evaluation ...
-
A Skill Library is an organized, governed collection of AI skills available across an enterprise — letting teams share, reuse, and compose AI capabilities rather than rebuilding from scratch. Matur...
-
A Skill Manager is the operational system that handles the lifecycle of AI skills — authoring, versioning, deployment, monitoring, retirement — across an enterprise AI portfolio. The role parallels...
-
A Skill Marketplace is a discoverable storefront where AI skills can be browsed, evaluated, and adopted — sometimes free, sometimes commercial. Public marketplaces include OpenAI's GPT Store (thous...
-
Skill Permissions control who can access which AI skills — enforcing role-based access, department boundaries, data sensitivity tiers, and regulatory requirements. A skill that can read patient rec...
-
A Skill Registry is the searchable directory of available AI skills — exposing skill metadata (name, description, input/output schemas, version, owner, tags, performance metrics) so applications an...
-
Skill Routing is the automated process of selecting the most appropriate skill (or model) to handle a given request — typically driven by an LLM analyzing the request and choosing from available op...
-
Skill Testing systematically evaluates AI skill behavior against expected outcomes — running automated test suites whenever a skill changes to prevent regression and validate improvements. Test cas...
-
Skill Versioning tracks every change to an AI skill — prompt updates, schema changes, knowledge-base refreshes, model substitutions — and enables rollback when a new version performs worse than the...
-
A Soft Prompt is a learned vector embedding that conditions a frozen language model toward a specific task — distinct from hard prompts which use discrete tokens (actual words). Soft prompts live i...
-
A Sparse Model is a neural network in which only a fraction of parameters or activations are active for any given input — contrasting with dense models where all parameters participate in every com...
-
Speaker Diarization is the task of identifying who spoke when in audio recordings — segmenting audio by speaker without necessarily knowing who any speaker is ("speaker 1," "speaker 2") or identify...
-
Speculative Decoding accelerates LLM inference by using a small "draft" model to propose several candidate tokens at once, which the larger "target" model then verifies in a single forward pass. Wh...
-
Speech Synthesis (also called Text-to-Speech or TTS) converts written text into natural-sounding spoken audio. The field has been transformed by neural approaches: classical concatenative and forma...
-
SPLADE (Sparse Lexical and Expansion Model for First Stage Ranking) is a sparse-retrieval model family that produces sparse, vocabulary-aligned vectors — combining the interpretability of tradition...
-
A Stop Sequence is a string that, when generated by an LLM, signals the model to immediately stop producing output — used to enforce response boundaries, prevent over-generation, and structure outp...
-
Streaming Inference returns AI output progressively as it is generated, rather than waiting for the complete response. For LLMs this means streaming tokens word by word — the experience users now e...
-
Streaming Output returns LLM responses progressively as tokens are generated — appearing word by word rather than waiting for the complete response. The pattern dramatically improves perceived resp...
-
Structured Output is AI-generated content that conforms to a specific machine-parseable format — typically JSON matching a schema, XML, CSV, or domain-specific languages like SQL. The pattern is fo...
-
Symbolic AI (also called Good Old-Fashioned AI or GOFAI) refers to the classical AI approach that dominated from the 1950s through the 1980s — manipulating symbolic representations of knowledge usi...
-
A Taxonomy is a hierarchical classification system that organizes concepts into nested categories — like the biological taxonomy organizing life into kingdoms, phyla, classes, orders, families, gen...
-
TensorRT is NVIDIA's high-performance deep learning inference SDK, optimized specifically for NVIDIA GPUs to deliver lowest-latency, highest-throughput inference. The toolkit applies optimizations ...
-
Test-Time Compute (also called inference-time compute or thinking compute) is the strategy of spending more compute at inference time to improve answer quality — typically by allowing the model to ...
-
text-embedding-3-large is OpenAI's premium embedding model released in January 2024 — producing higher-quality vector representations than the smaller variant at higher cost. The model produces 307...
-
text-embedding-3-small is OpenAI's small embedding model released in January 2024 — designed as a high-performance, low-cost vector embedding option for production applications. The model produces ...
-
text-embedding-ada-002 is OpenAI's legacy embedding model from December 2022 — the model that defined commercial embedding for nearly a year and a half before being superseded by the text-embedding...
-
Text-to-Speech (TTS) is the technology that produces spoken audio from written text — synonymous with speech synthesis but emphasizing the input-output direction. Modern TTS quality is so high that...
-
A Token Counter measures how many tokens a piece of text uses according to a specific tokenizer — essential for predicting LLM costs, managing context-window limits, and optimizing prompts. Differe...
-
A Tool Registry is the catalog of external tools (functions, APIs, services) that AI agents can invoke — making capabilities discoverable, governable, and reusable. Registries store tool metadata: ...
-
Topic Modeling discovers thematic patterns across large document collections — typically without supervision — surfacing the recurring topics that organize the content. Classical algorithms include...
-
Tree of Thoughts (ToT) is a reasoning technique that explores multiple solution paths in parallel, evaluates them, and selects the most promising — rather than committing to a single chain of thoug...
-
Universal Sentence Encoder (USE) is Google Research's embedding model family released in 2018 — one of the first widely-used sentence-embedding models that produced strong semantic representations ...
-
Visual Question Answering (VQA) is the task of answering natural-language questions about images — "What color is the car?", "How many people are in the photo?", "What's the chart's main conclusion...
-
vLLM is an open-source LLM inference engine developed at UC Berkeley that became one of the most popular high-throughput serving systems for large language models. Its hallmark innovation is PagedA...
-
Voice Cloning produces synthetic speech that mimics a specific person's voice — using neural TTS models trained or conditioned on samples of the target voice. The technology has rapidly improved: m...
-
Voyage AI is a specialized embedding-model provider founded by AI researchers from Stanford — focused exclusively on producing the highest-quality embedding models for retrieval applications. The V...
-
A Web Agent is an AI system that operates on the web — browsing, searching, retrieving, summarizing, and acting on web content. Web agents are a broader category than browser agents (which specific...
-
Whisper is OpenAI's open-weight speech recognition model released in September 2022 — trained on 680,000 hours of multilingual and multitask audio from the web. The model supports automatic speech ...
-
An Activation Function introduces non-linearity into a neural network, allowing it to learn complex patterns rather than just linear relationships. Without activation functions, even a hundred-laye...
-
Active Learning lets a model request labels for the data points it finds most informative, dramatically reducing labeling cost. Instead of randomly labeling thousands of examples, the model identif...
-
An Adversarial Attack is a deliberate attempt to manipulate an AI system through crafted inputs designed to cause incorrect or unintended behavior. Famous adversarial attacks include the demonstrat...
-
An Adversarial Example is an input crafted to fool an AI model — often by introducing perturbations imperceptible to humans but devastating to model accuracy. The phenomenon was first widely docume...
-
Agent Memory is the persistent storage that lets an AI agent recall past interactions, learnings, or user context across sessions — so the agent does not start every conversation as a stranger. Mem...
-
Agentic AI refers to systems where language models plan, reason, and act with significant autonomy — chaining tools, calling APIs, and pursuing multi-step goals without step-by-step human direction...
-
AI Accountability is the principle that humans — not algorithms — bear responsibility for the outcomes of AI systems. It requires clear chains of responsibility, decision rights, and consequences f...
-
An AI Agent is a system that perceives its environment, makes decisions, and takes actions to achieve goals — often calling tools or APIs autonomously. Modern AI agents combine an LLM brain with a ...
-
AI Alignment is the technical and philosophical challenge of making AI systems pursue goals that match human values and intentions — including goals their designers didn't anticipate but would endo...
-
An AI Approval Workflow is a structured process that routes proposed AI use cases through legal, security, AI ethics, and business approvers before development or deployment. Typical workflows incl...
-
An AI Assistant is a conversational AI system that helps users accomplish tasks via natural language. Consumer examples include OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini, Apple Intellig...
-
An AI Audit Trail is a complete, tamper-evident record of every AI interaction — including who used the AI, when, what they asked, what context was retrieved, what the model produced, and what acti...
-
The Blueprint for an AI Bill of Rights, published by the White House Office of Science and Technology Policy in October 2022, is a non-binding U.S. policy document articulating five principles for ...
-
An AI Center of Excellence (CoE) is a centralized team that sets standards, builds reusable platforms, and supports business-unit AI delivery across an enterprise. The CoE typically owns AI strateg...
-
AI Compliance is the operational practice of meeting legal, regulatory, and contractual obligations applicable to AI systems — including evidence collection, documentation, audit response, and ongo...
-
AI Conformity is the state of an AI system meeting the requirements of applicable laws, standards, and contractual obligations. The term is used broadly to encompass conformity assessment under the...
-
An AI Copilot is an AI assistant embedded inside a tool — a code editor, document app, CRM, or design platform — that supports the user without taking full control. The term was popularized by GitH...
-
AI Discrimination occurs when an AI system produces decisions that disadvantage protected groups in ways prohibited by law, contract, or AI ethics policy. The concept extends traditional anti-discr...
-
AI Documentation is the collection of artifacts describing an AI system — model cards, datasheets for datasets, system architecture diagrams, evaluation reports, risk assessments, monitoring runboo...
-
AI Ethics is the field that examines how AI systems should be designed and deployed to respect human values, rights, and well-being. Core principles across most frameworks include fairness, account...
-
An AI Ethics Board is a cross-functional group of internal and external stakeholders that reviews high-risk AI use cases and advises leadership on AI ethics and AI policy. Boards typically include ...
-
AI Fairness is the discipline of ensuring AI systems treat individuals and groups equitably across protected and relevant characteristics. Fairness is not one thing — researchers have catalogued do...
-
AI Governance is the framework of policies, roles, controls, and processes an organization uses to manage AI responsibly across its lifecycle — from idea to retirement. Mature AI governance include...
-
An AI Impact Assessment is a structured evaluation of an AI system's potential effects on individuals, groups, and society — covering fairness, privacy, security, safety, and broader impacts. Often...
-
An AI Incident is an event in which AI behavior causes — or could have caused — harm, financial loss, regulatory violation, or significant operational disruption. Real-world AI incidents tracked in...
-
The AI Lifecycle describes the stages an AI system passes through: ideation, design, data collection, training, validation, deployment, monitoring, iteration, and retirement. Each stage involves sp...
-
AI Lifecycle Management governs every stage of an AI system's life — from ideation and design through development, validation, deployment, monitoring, and eventual retirement. Each stage has distin...
-
An AI Maturity Model assesses how advanced an organization's AI capabilities are across dimensions like strategy, data, talent, infrastructure, governance, and ethics — typically scoring each on a ...
-
An AI Notified Body is a third-party organization designated by an EU member state to perform conformity assessments for certain high-risk AI systems under the EU AI Act. Notified Bodies must demon...
-
The AI Operating Model defines how AI is organized, funded, governed, and delivered inside a company — the org chart, roles, decision rights, and processes of an enterprise AI function. Common patt...
-
AI Orchestration coordinates multiple models, tools, prompts, and data sources to deliver complex enterprise AI workflows. Rather than a single LLM call, orchestration platforms route requests thro...
-
AI Penetration Testing extends traditional security pen testing to AI-specific attack surfaces. While AI red teaming focuses broadly on undesirable behavior including bias and safety, AI pen testin...
-
An AI Plugin extends a language model with access to external capabilities — search, calculations, databases, browsing, image generation, or specialized APIs. The category was popularized by ChatGP...
-
AI Policy is a written organizational guideline that defines acceptable AI use, required controls, and roles and responsibilities for everyone interacting with AI systems. Typical AI policies cover...
-
AI Red Teaming is the practice of probing AI systems for vulnerabilities, biases, jailbreaks, prompt-injection susceptibility, and unsafe behaviors before adversaries find them in production. Red t...
-
AI Risk Management identifies, assesses, mitigates, and monitors risks specific to AI systems across the lifecycle. Risk categories include AI-specific concerns (bias, hallucination, prompt injecti...
-
An AI Risk Register is a living catalog of identified AI risks across the enterprise — each entry capturing risk description, likelihood, impact, mitigation status, and owner. Common risk categorie...
-
AI Safety is the field focused on preventing AI systems from causing harm — through accidents, misuse, misalignment, or unforeseen capabilities. AI safety operates at multiple levels: technical saf...
-
AI Stewardship assigns clear ownership and accountability for each AI system to a specific person or team responsible for its design, deployment, and lifecycle outcomes. The principle borrows from ...
-
AI Strategy is the long-term plan that aligns AI investments with business objectives, technical capabilities, and risk tolerance. A mature AI strategy answers fundamental questions: which use case...
-
An AI System Inventory is a comprehensive registry of every AI system an organization builds, buys, or uses — including third-party APIs, embedded vendor AI, and internal models. The inventory typi...
-
AI Transparency is the practice of making AI systems' purpose, capabilities, limitations, training data, and behaviors visible to relevant stakeholders. Transparency operates at multiple levels: to...
-
An AI Use Case Registry catalogs every AI application across an organization — what it does, who owns it, what risks it carries, and what controls apply. Each use case is typically registered befor...
-
An AI Vulnerability is a weakness in an AI system that can be exploited by attackers to cause unauthorized behavior, data leakage, or harm. Categories include prompt injection (manipulating model b...
-
AIOps (Artificial Intelligence for IT Operations) applies AI to IT infrastructure management — including event correlation, anomaly detection, predictive maintenance, and root-cause analysis across...
-
An Algorithm is a precise, step-by-step procedure for solving a problem or producing an output. In AI, algorithms range from simple decision trees and linear regression to sophisticated deep neural...
-
The Algorithmic Accountability Act is proposed U.S. federal legislation (versions introduced 2019, 2022, and 2023) that would require companies to conduct impact assessments of automated decision s...
-
Algorithmic Bias is systematic, repeatable unfairness in an AI system's output that disadvantages certain groups — often along lines of race, gender, age, disability, or socioeconomic status. Famou...
-
Approximate Nearest Neighbor (ANN) search finds the closest matches in a vector space without examining every candidate — trading a small amount of accuracy for enormous speed gains. Exact nearest ...
-
Artificial Intelligence (AI) is the field of computer science focused on building systems that can perform tasks normally requiring human intelligence — reasoning, learning, perception, decision-ma...
-
An Artificial Neural Network (ANN) is the formal term for the layered, parameterized models that power deep learning. The name distinguishes them from biological neural networks, though the underly...
-
An Attention Mechanism lets a neural network focus on the most relevant parts of its input when producing each output, rather than treating all inputs equally. The concept was introduced for machin...
-
An Autonomous Agent is an AI system that operates with minimal human intervention to complete tasks. Autonomy exists on a spectrum: a low-autonomy agent might suggest actions for human approval; a ...
-
Backpropagation is the algorithm that computes how each parameter in a neural network contributed to error, enabling gradient descent to update those parameters intelligently. Popularized in 1986 b...
-
Batch Normalization stabilizes and accelerates neural network training by normalizing the inputs to each layer using statistics computed over the current mini-batch. Introduced by Ioffe and Szegedy...
-
A Bias Audit is a systematic review of an AI system for discriminatory outcomes across demographic groups, often required by law, contract, or organizational AI policy. New York City Local Law 144,...
-
Bias Mitigation comprises techniques applied at different stages of the AI lifecycle to reduce unfair outcomes. Pre-processing approaches modify training data — rebalancing, reweighting, or generat...
-
The Bias-Variance Tradeoff describes the tension between models that are too simple (high bias, underfitting) and those that are too complex (high variance, overfitting). High bias means the model ...
-
The California Consumer Privacy Act (CCPA), effective 2020 and expanded by the CPRA in 2023, is California's comprehensive privacy law. It gives California consumers rights to know what personal in...
-
CE Marking for AI is the mark of conformity affixed to high-risk AI systems that have successfully completed an EU AI Act conformity assessment, allowing them to be sold within the European Economi...
-
Chain-of-Thought (CoT) Prompting is a technique where a language model is asked to reason step-by-step before giving a final answer, dramatically improving accuracy on complex problems. The breakth...
-
Chunking splits long documents into smaller, retrievable segments for use in embeddings databases and retrieval-augmented generation pipelines. Chunk size matters enormously — too small and the mod...
-
Classification is a machine learning task that assigns inputs to discrete categories — fraud or not fraud, spam or ham, malignant or benign, or one of dozens of product categories. It is one of the...
-
Clustering groups similar data points together without labels — a form of unsupervised learning useful for customer segmentation, anomaly detection, and exploratory analytics. Popular algorithms in...
-
Concept Drift occurs when the relationship between inputs and outputs changes over time — meaning the same input now corresponds to a different output than it did at training time. While data drift...
-
Confirmation Bias is the human tendency to favor information that confirms existing beliefs — a problem that can creep into AI development at every stage. Engineers may unconsciously design experim...
-
A Conformity Assessment is a structured evaluation of whether an AI system meets specified regulatory or standards requirements. Under the EU AI Act, providers of high-risk AI systems must complete...
-
Constitutional AI (CAI) is an alignment technique developed by Anthropic where AI models are trained against a written set of principles — a "constitution" — that they use to critique and revise th...
-
Context Augmentation enriches a language model's input with retrieved content, structured data, or computed results before generation — the broader pattern that includes RAG and other techniques. A...
-
The Context Window is the maximum amount of text — measured in tokens — that an AI model can process at once. It includes the system prompt, conversation history, retrieved documents, and the lates...
-
A Convolutional Neural Network (CNN) is a specialized neural network architecture that excels at processing images, video, and other grid-like data. CNNs use convolutional layers that scan small fi...
-
Cosine Similarity measures the angle between two vectors, producing a value between -1 and 1 that captures how similar two embeddings are in direction (regardless of magnitude). It is the most comm...
-
Cross-Validation rotates training and validation splits across the dataset to produce more robust performance estimates than a single split would provide. The most common form, k-fold cross-validat...
-
Data Bias is unfair representation in training data that causes AI systems to perform unevenly across groups or contexts. Sources include historical data reflecting past discrimination (lending rec...
-
Data Drift is a change in the statistical properties of model inputs after deployment — meaning the data the model sees in production differs from the data it was trained on. Common forms include f...
-
Data Lineage is the end-to-end trace of data flow through systems — from origin through every transformation, join, aggregation, and use. While related to provenance (which focuses on origin and hi...
-
Data Minimization is the principle of using the least amount of personal data necessary to accomplish a stated purpose — a foundational requirement under GDPR, CCPA, and most modern privacy laws. I...
-
Data Poisoning is an attack in which an adversary corrupts training data to manipulate the resulting AI model's behavior. The attack might insert specific examples that cause the model to misclassi...
-
Data Provenance is the documented history of where data came from, how it was collected, what transformations it underwent, and how it has been used. Strong provenance answers questions like: who c...
-
A Datasheet for Datasets is a structured document — modeled on hardware spec sheets — that describes a dataset's purpose, composition, collection process, preprocessing, recommended uses, and limit...
-
Deep Learning is a subset of machine learning that uses multi-layered neural networks to model complex patterns in data such as images, audio, video, and text. The field took off after 2012 when Al...
-
Demographic Parity is a fairness criterion requiring that an AI system produce positive outcomes at equal rates across demographic groups, regardless of underlying differences in the population. Fo...
-
Differential Privacy is a mathematical framework for providing strong privacy guarantees while still extracting useful statistics from data. By adding carefully calibrated noise to queries or train...
-
A Diffusion Model generates content by gradually denoising random input over many small steps, learning to reverse a process of progressively adding noise to real data. The approach, popularized by...
-
Dimensionality Reduction compresses many input variables into fewer meaningful ones using techniques like Principal Component Analysis (PCA), t-SNE, UMAP, and autoencoders. The goal is to retain th...
-
Disparate Impact is a legal and statistical concept that captures when a facially neutral policy or AI system produces significantly unequal outcomes across protected groups. The concept originated...
-
Document Indexing transforms raw documents — PDFs, Word files, HTML pages, transcripts, spreadsheets — into a searchable, retrievable knowledge base for AI systems. The process typically involves p...
-
Dropout is a regularization technique that randomly disables (or "drops out") a fraction of neurons during training to reduce overfitting. Introduced by Srivastava, Hinton, and colleagues in 2014, ...
-
An Embedding is a dense vector representation of data — like words, sentences, images, or audio — that captures semantic meaning in a fixed-length array of numbers (typically 256 to 4,096 dimension...
-
An Embeddings Database is a specialized store for vector representations of content, enabling fast semantic lookup at scale. The term overlaps heavily with vector database, though some practitioner...
-
An Encoder-Decoder architecture is a neural network design where one component compresses input into a representation and another generates output from it. The encoder reads the entire input (a sen...
-
Equalized Odds is a fairness metric requiring that an AI's error rates — both false positives and false negatives — be equal across protected groups. Formalized by Hardt, Price, and Srebro in 2016,...
-
The EU AI Act is the European Union's comprehensive regulation of artificial intelligence — the world's first major horizontal AI law. Adopted in 2024 with phased enforcement through 2026 and beyon...
-
The U.S. Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence (EO 14110) was issued by President Biden in October 2023 and partially rescinded by the ...
-
Explainable AI (XAI) provides understandable reasons for AI predictions or actions, making decisions inspectable for users, auditors, and regulators. Techniques include feature importance (SHAP, LI...
-
A Fairness Metric is a quantitative measure of how equitably an AI system performs across groups. Different metrics capture different conceptions of fairness — demographic parity, equalized odds, p...
-
A Feature is an individual input variable used by an AI model — like age, transaction amount, pixel intensity, or word frequency. Feature selection and engineering directly influence fairness, accu...
-
Federated Learning is a privacy-preserving technique that trains AI models across many devices or organizations without centralizing raw data. Each participant trains a local copy of the model on i...
-
Few-Shot Learning lets an AI model perform a new task after seeing only a handful of examples, often provided directly in the prompt as demonstrations. The technique was popularized by the GPT-3 pa...
-
Fine-Tuning adapts a pretrained AI model to a specific task or domain using a smaller, targeted dataset — typically dozens to thousands of examples rather than the billions used in pretraining. It ...
-
A Foundation Model is a large, general-purpose AI model trained on broad data and adapted to many downstream tasks through prompting, fine-tuning, or other adaptation techniques. The term was popul...
-
Function Calling lets a language model invoke structured tools or APIs with typed arguments, returning the function's output back into the conversation. The pattern was popularized by OpenAI's June...
-
A Gated Recurrent Unit (GRU), introduced by Cho et al. in 2014, is a streamlined alternative to LSTM that combines the forget and input gates into a single update gate. With fewer parameters than a...
-
The General Data Protection Regulation (GDPR) is the European Union's comprehensive privacy law, effective since May 2018 and binding on any organization processing EU residents' personal data. Key...
-
A Generative Adversarial Network (GAN) pairs two neural networks — a generator and a discriminator — that compete during training. The generator produces synthetic data trying to fool the discrimin...
-
Generative AI refers to systems that create new content — text, images, code, audio, video, or 3D models — rather than just classifying or predicting from existing inputs. The category exploded int...
-
Gradient Descent is the optimization algorithm that drives most modern AI training — iteratively adjusting model parameters to reduce error. The basic idea is intuitive: compute the slope (gradient...
-
Grounding constrains a generative AI model to verified sources of truth, dramatically reducing hallucination. The most common technique is retrieval-augmented generation (RAG), where the system ret...
-
Hallucination is when a generative AI model produces confident but false information — invented facts, fake citations, imaginary entities, or fabricated quotations. Famous examples include a New Yo...
-
A High-Risk AI System is an AI application that the EU AI Act subjects to its strictest obligations because of significant potential for harm to health, safety, or fundamental rights. Annex III of ...
-
The Health Insurance Portability and Accountability Act (HIPAA) is the U.S. federal law governing protected health information (PHI) in healthcare and health-adjacent contexts. The Privacy Rule, Se...
-
Human Oversight ensures qualified people supervise AI systems — with clear authority, training, and tools to intervene when necessary. Oversight differs from human-in-the-loop in that it can includ...
-
Human-in-the-Loop (HITL) keeps humans involved in AI decision-making — reviewing, approving, or correcting AI outputs before they affect the real world. The pattern is essential in high-stakes doma...
-
Hybrid Search combines lexical (keyword-based) and semantic (vector-based) retrieval to deliver more accurate results than either approach alone. The technique runs both searches in parallel and me...
-
A Hyperparameter is a configuration value set before training — like learning rate, tree depth, dropout rate, or number of hidden layers — that controls how a model learns rather than what it learn...
-
Instruction Tuning is a fine-tuning technique that teaches a language model to follow natural-language instructions — turning a raw foundation model into a useful assistant. The base pretrained mod...
-
Interpretability is the property of an AI model that allows humans to understand how it makes decisions. While related to explainability, interpretability typically refers to deeper, mechanistic un...
-
ISO/IEC 23894 is an international standard providing guidance on managing risks specific to AI systems, published in February 2023. It complements ISO 31000 (general risk management) and ISO/IEC 42...
-
ISO/IEC 42001 is the first international standard for AI management systems, published in December 2023. Modeled after ISO/IEC 27001 (information security management), it provides a certifiable fra...
-
A Knowledge Base is a curated collection of information — articles, FAQs, policies, manuals, troubleshooting guides — that AI systems can retrieve from to provide grounded answers. Traditional know...
-
A Knowledge Graph is a structured representation of entities (people, products, concepts) and the relationships between them, used to enrich AI reasoning and retrieval. Famous knowledge graphs incl...
-
A Label is the correct answer attached to a training example in supervised machine learning — like "fraud" or "not fraud," "cat" or "dog," or a numerical value like "price = $245." Labels can be pr...
-
A Large Language Model (LLM) is a transformer-based AI model trained on massive text corpora to generate, summarize, translate, and reason over language. Modern LLMs contain billions to trillions o...
-
Layer Normalization is an alternative to batch normalization that normalizes across the features within each individual example rather than across the batch dimension. Introduced by Ba, Kiros, and ...
-
Lexical Search retrieves information using exact word matches or close variations, with classical algorithms like BM25, TF-IDF, and inverted indexes. Lexical search is the traditional approach behi...
-
LLMOps applies operational discipline to large language models — adding capabilities like prompt versioning, evaluation suites, RAG pipeline observability, token-level cost tracking, and content-sa...
-
Long Short-Term Memory (LSTM) is a type of recurrent neural network designed by Hochreiter and Schmidhuber in 1997 to retain information over long sequences by using gating mechanisms (input, forge...
-
Machine Learning (ML) is a branch of AI in which systems learn patterns from data rather than being explicitly programmed. The term was coined by Arthur Samuel at IBM in 1959 and has grown into the...
-
MLOps (Machine Learning Operations) is the discipline of automating, monitoring, and reliably operating machine learning systems in production. Inspired by DevOps, MLOps adds practices specific to ...
-
An AI Model is the trained output of a machine learning process — the mathematical artifact that turns inputs into predictions, classifications, or generated content. Models can be as simple as a l...
-
A Model Card is a standardized document that describes a machine learning model's purpose, training data, intended uses, limitations, evaluation results, and known risks. The concept was introduced...
-
Model Context Protocol (MCP) is an open standard introduced by Anthropic in late 2024 that lets AI assistants connect to tools, data sources, and services in a consistent, vendor-neutral way. Rathe...
-
Model Drift is the degradation of an AI model's performance over time as the world changes around it — even when the model code stays the same. Drift comes from many causes: customer behavior shift...
-
Model Monitoring is the continuous observation of deployed AI systems to detect performance degradation, drift, anomalies, and operational issues. A complete monitoring program covers technical met...
-
Model Risk is the potential for adverse outcomes from errors, limitations, or misuse of AI and machine-learning models. The discipline originated in financial services with the Federal Reserve's SR...
-
Model Validation is the independent verification that an AI model performs as intended before and after deployment. Borrowed from financial services model risk management (Federal Reserve SR 11-7),...
-
Model Verification confirms that an AI model is built correctly — that it implements its intended specification, behaves consistently across environments, and meets defined technical requirements. ...
-
A Multi-Agent System orchestrates several AI agents that collaborate, negotiate, or compete to solve complex tasks. Each agent might specialize in a different role — a research agent, a writer agen...
-
Multi-Head Attention runs several self-attention computations in parallel, allowing transformers to capture different types of relationships simultaneously — one head might focus on syntactic struc...
-
Multimodal AI processes and generates more than one type of data — combining text, images, audio, video, and sometimes 3D or sensor data. Modern multimodal models include OpenAI's GPT-4o (text + im...
-
A Neural Network is a computational model loosely inspired by the brain, made of interconnected layers of nodes ("neurons") that learn from data by adjusting the strengths of their connections. Eac...
-
The NIST AI Risk Management Framework (AI RMF 1.0, released January 2023) is a voluntary U.S. government framework for managing risks of AI systems. Built around four functions — Govern, Map, Measu...
-
Online Learning continuously updates an AI model as new data arrives, instead of training in batches. This makes it ideal for environments where data patterns shift rapidly and waiting hours or day...
-
Operational Risk in AI is the risk of loss from inadequate or failed internal processes, people, and systems supporting AI — or from external events disrupting AI operations. Examples include AI se...
-
Overfitting happens when an AI model memorizes its training data instead of learning generalizable patterns, leading to excellent training performance but poor real-world results. A classic example...
-
Personally Identifiable Information (PII) is data that can identify a specific individual — either alone or in combination with other available data. Direct identifiers include name, Social Securit...
-
Pretraining is the initial, large-scale training phase where a foundation model learns general patterns from massive datasets before being adapted to specific tasks. For a modern LLM, pretraining m...
-
Prohibited AI Practices are AI uses banned outright under the EU AI Act because of unacceptable risk. The list includes manipulative AI that exploits vulnerabilities to cause harm, social scoring b...
-
A Prompt is the input text given to a generative AI model to produce a response. Prompts shape model behavior more than most users realize — small changes in wording, formatting, or context can dra...
-
Prompt Engineering is the discipline of crafting effective inputs to large language models to achieve reliable, safe, and high-quality outputs. Techniques include zero-shot prompting (just ask), fe...
-
Prompt Injection is a security attack in which malicious instructions are hidden in inputs to trick a language model into bypassing its safety guardrails, leaking confidential data, or performing u...
-
ReAct (Reasoning + Acting) is an agent design pattern introduced by Yao et al. (2022) where a language model alternates between reasoning steps ("Thought:") and tool actions ("Action:") to solve co...
-
A Recurrent Neural Network (RNN) processes sequences such as text, audio, or time-series data by maintaining a hidden state that carries information from one step to the next. This makes RNNs natur...
-
Regression is a machine learning task that predicts continuous numeric values — like price, demand, temperature, or risk score — rather than discrete categories. Linear regression is the simplest f...
-
A Regulatory Sandbox lets companies test innovative AI applications under regulator supervision with relaxed rules, helping regulators learn while supporting innovation. The EU AI Act mandates that...
-
Reinforcement Learning (RL) trains agents to make decisions by rewarding desirable behavior and penalizing mistakes — much like teaching a dog with treats. An RL agent interacts with an environment...
-
ReLU (Rectified Linear Unit) is the most widely used activation function in modern deep learning. Its definition is dead simple — output zero for negative inputs and the input itself for positive v...
-
Reputational Risk is the potential damage to brand, customer trust, and stakeholder relationships from AI failures or controversies. Examples that made headlines include the 2016 Microsoft Tay chat...
-
Reranking is a second-pass retrieval step that re-orders an initial set of candidate documents using a more expensive but more accurate model — typically a cross-encoder that examines query and doc...
-
Responsible AI is the umbrella discipline encompassing AI ethics, fairness, transparency, accountability, privacy, security, safety, and AI compliance — applied across the AI lifecycle from concept...
-
Retrieval-Augmented Generation (RAG) combines a language model with a search system that fetches relevant documents before answering, dramatically reducing hallucination and letting the model use k...
-
RLHF (Reinforcement Learning from Human Feedback) uses human preferences to align large language models with desired behaviors. The process trains a reward model on human comparisons ("which of the...
-
Sampling Bias arises when training data fails to represent the population the AI will later serve. The classic example is the 1936 Literary Digest poll, which predicted Alf Landon would defeat FDR ...
-
Selection Bias occurs when the process by which data is collected systematically distorts outcomes — affecting validity of any AI trained on that data. Classic forms include survivorship bias (only...
-
Self-Attention is a specific attention pattern where every element of a sequence relates to every other element, enabling transformers to model rich context within a single input. For each token in...
-
Self-Supervised Learning teaches models to generate their own labels from raw data — for example, predicting the next word in a sentence or filling in a masked image patch. By turning unlabeled dat...
-
Semantic Search retrieves information based on meaning rather than exact keyword match, using embeddings and vector similarity. Unlike traditional keyword search where you must guess the exact term...
-
Semi-Supervised Learning blends a small set of labeled data with a large pool of unlabeled data to train more capable models cost-effectively. The technique exploits the fact that unlabeled data is...
-
Shadow AI is unauthorized AI use within an organization — employees using public LLMs for work tasks without IT approval, business units deploying AI tools without governance review, or embedded AI...
-
The Sigmoid function squashes any input into a value between 0 and 1 using the formula 1/(1+e^-x). The output looks like an S-curve, smoothly transitioning from near-zero for large negative inputs ...
-
SOC 2 (Service Organization Control 2) is a widely adopted compliance framework developed by the AICPA that assesses how service organizations protect customer data across five Trust Services Crite...
-
Softmax is an activation function that converts a vector of numbers into a probability distribution — every output is between 0 and 1, and they all sum to 1. It is the standard final layer in multi...
-
Speech-to-Text (STT) AI converts spoken audio into written text — also called Automatic Speech Recognition (ASR). The field has matured dramatically with models like OpenAI's Whisper (open weights,...
-
Supervised Learning is a machine learning approach where models learn from labeled examples — each input is paired with a known correct output. The model learns to predict outputs for new inputs by...
-
Synthetic Data is artificially-generated data that imitates the statistical properties of real data without revealing real individuals or sensitive records. Generation techniques include GANs, diff...
-
A System Prompt is a hidden, foundational instruction given to a language model that defines its persona, rules, and constraints before any user input. The system prompt typically establishes who t...
-
Task Decomposition is the process by which an AI agent breaks a complex goal into smaller, executable steps. Given a high-level objective like "plan a four-day trip to Tokyo for a family of four wi...
-
Temperature is a generative AI setting that controls output randomness — lower values produce focused, deterministic responses; higher values produce more creative, varied ones. At temperature 0, t...
-
Test Data is held out from training and used to evaluate how an AI model performs on unseen examples. The strict separation between training and test data is one of the oldest disciplines in machin...
-
Text-to-Image AI generates images from natural-language descriptions using diffusion or transformer models. The category exploded with the public release of DALL-E 2, Stable Diffusion, and Midjourn...
-
Text-to-Video AI generates moving images from natural-language prompts, an emerging frontier of generative AI that builds on text-to-image and diffusion model breakthroughs. OpenAI's Sora, demonstr...
-
Third-Party AI Risk is the risk that AI capabilities sourced from external vendors, partners, or embedded inside other software cause problems for the buyer. Examples include vendor service outages...
-
A Token is the smallest unit of input an AI language model processes — typically a word, sub-word, or character depending on the tokenizer. As a rough rule of thumb in English, one token equals abo...
-
Tokenization is the process of breaking text into smaller units (tokens) that an AI model can process. Modern large language models use subword tokenizers like Byte-Pair Encoding (BPE), WordPiece, ...
-
Tool Use is the ability of an AI model — typically an agent — to call external functions, APIs, or services to extend its capabilities beyond text generation. Common tools include search engines (G...
-
Top-K Sampling restricts a generative AI model to choosing among its K most likely next tokens, providing a different tradeoff between diversity and reliability than top-P. With top_k = 50, the mod...
-
Top-P Sampling (nucleus sampling) is a generative AI technique that limits a model's word choices to the smallest set of probable tokens whose cumulative probability exceeds threshold P. For exampl...
-
Training Data is the labeled or unlabeled information used to teach an AI model. Its quality, representativeness, and provenance shape every downstream behavior — including bias, accuracy, and the ...
-
Transfer Learning takes a model trained on one task and adapts it to a new, related task with much less data — a cornerstone of modern enterprise AI. Instead of training a giant model from scratch ...
-
The Transformer is the neural network architecture that revolutionized AI, introduced in the 2017 paper "Attention is All You Need" by Vaswani et al. at Google. Transformers replaced recurrence wit...
-
Trustworthy AI is a term used by the EU, NIST, and other authorities to describe AI systems that are lawful, ethical, and robust — operating reliably while respecting human rights and democratic va...
-
Underfitting occurs when an AI model is too simple to capture the patterns in its data, producing weak predictions on both training and test sets. A classic example is trying to fit a straight line...
-
Unsupervised Learning is a class of machine learning algorithms that finds hidden patterns in data without labeled outputs. Instead of being told the right answer, the model discovers structure on ...
-
Validation Data is used during model development to tune hyperparameters and select the best version of an AI model — separate from both training and test data. A typical split might be 70% trainin...
-
In AI, a Vector is an ordered list of numbers that represents data in a way machines can process. A 768-dimensional vector might encode a sentence's meaning; a 2,048-dimensional vector might repres...
-
A Vector Database stores embeddings and supports fast similarity search across billions of vectors, powering RAG, semantic search, recommendation systems, and image retrieval. Major options include...
-
A Vector Index is the data structure inside a vector database that enables fast similarity search across millions or billions of high-dimensional embeddings. Common index algorithms include HNSW (H...
-
A Word Embedding is a vector representation of a word that captures its meaning relative to other words in the vocabulary. The breakthrough came with Word2Vec (Mikolov et al., 2013) and GloVe (Penn...
-
Workflow Automation uses AI — increasingly agentic AI — to execute multi-step business processes end to end, from triggering events through completion. Traditional workflow tools (Zapier, n8n, Micr...
-
Zero-Shot Learning is the ability of a model to perform a task without any task-specific examples — relying entirely on its pretraining knowledge and the instruction given in the prompt. This capab...