Vector Index
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 vector among millions would require comparing the query against every stored vector — an O(n) operation that quickly becomes infeasible. Common vector index types include HNSW (graph-based), IVF (cluster-based), IVF-PQ (cluster plus compression), LSH (hashing), and DiskANN (disk-resident graphs), each making different trade-offs between recall, latency, memory footprint, and build time. Choosing the right index for a workload — and tuning parameters like efConstruction, M, and nprobe — has dramatic effects on both performance and cost. AI governance frameworks pay attention to vector index configuration because under-tuned recall can silently degrade
RAG answer quality, while over-tuned recall wastes infrastructure budget. Most production deployments validate index quality with Recall@k benchmarks against a ground-truth subset before going live.
Vector index choice with Centralpoint: Centralpoint stays model-agnostic across whatever vector index you operate — HNSW in Weaviate, IVF-PQ in Milvus, pgvector with HNSW — and meters the retrieval-plus-generation token cost so finance sees the real economics. Prompts and skill definitions stay on-premise while chatbots that depend on vector lookups embed via one line of JavaScript.
Related Keywords:
Vector Index,
,