Chunk Overlap

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 retrieval. Typical overlap values are 10-20% of chunk size — for 500-token chunks, 50-100 tokens of overlap is common. Overlap preserves context across boundary regions: a sentence spanning two chunks appears (at least partially) in both, increasing the chance that either chunk can be retrieved for queries that reference that sentence. The trade-off is storage and retrieval cost — overlap multiplies the number of vectors and the amount of duplicate content the vector index must store. Higher overlap also increases the chance that retrieval returns multiple chunks containing the same information, requiring deduplication in post-retrieval processing. AI governance teams document chunk overlap alongside chunk size as part of their RAG architecture lineage. Frameworks like LangChain default to 10-20% overlap, while domain-specific implementations sometimes use higher (30-50%) for legal or medical content where boundary loss is unacceptable.

Chunk overlap configuration in Centralpoint: Centralpoint sits above whatever chunking pipeline you operate, metering tokens across the resulting retrievals so the cost-quality trade-off of overlap is transparent. The model-agnostic platform keeps prompts on-premise, supports both generative and embedded models, and deploys chatbots through one line of JavaScript.


Related Keywords:
Chunk Overlap,,