Chunk Overlap
ProcessingChunk overlap is the amount of content shared between adjacent chunks. For example, if chunks are 500 tokens with 50-token overlap, the last 50 tokens of chunk N also appear at the start of chunk N+1. Overlap can improve recall when relevant content falls near chunk boundaries.
Trade-offs
- Recall — Overlap reduces the chance that key content is split across chunks and missed by retrieval
- Redundancy — Overlapping chunks increase index size and can return duplicate context for the LLM
- Cost — More overlapping chunks mean more embeddings and storage
DocLD uses semantic chunking by default, which respects logical boundaries. Knowledge bases can be configured with different chunking settings, including overlap.
Related Concepts
Chunk overlap is a chunking strategy that affects embedding and vector search. Top-k and reranking can mitigate recall issues without overlap, but overlap remains a common option for retrieval-heavy use cases.