What is Chunk Overlap?
Chunk Overlap is the repeated text shared between adjacent document chunks so that information near a split boundary remains available during retrieval.
How It Works
Chunk overlap is a practical safeguard against cutting a sentence, definition, table explanation, or procedure across two chunks. It can improve recall for questions whose answer sits near a boundary, but excessive overlap inflates the index, increases duplicated evidence, and can crowd out diverse results. Good overlap policies are usually modest, token-based, and paired with deduplication or reranking so the generator does not see the same evidence repeatedly.
Key Characteristics
- Preserves local context around chunk boundaries
- Often configured as a token count or percentage of chunk size
- Improves recall for boundary-spanning facts but increases index size
- Can create near-duplicate search results if top-k retrieval is not diversified
- Works best when combined with structure-aware splitting and reranking
Common Use Cases
- Keeping a definition and its following explanation together across a split
- Reducing answer failures when paragraphs are longer than the target chunk size
- Maintaining code comments and code lines near a chunk boundary
- Testing overlap sizes during RAG retrieval evaluation
- Using overlap selectively for prose while avoiding duplication in structured records
Example
Loading code...Frequently Asked Questions
How much chunk overlap should a RAG system use?
A modest overlap such as a small fraction of the chunk size is a common starting point, but the right value depends on document structure, query type, and duplicate-result tolerance.
Can chunk overlap hurt retrieval?
Yes. Too much overlap creates duplicate vectors and repeated results, which can reduce diversity and waste context-window space.
Is overlap needed for structured data?
Often less so. Records, rows, and well-bounded fields usually need metadata and schema preservation more than sliding text overlap.
How does overlap affect citations?
Overlap can preserve evidence near boundaries, but systems should keep source offsets so citations still point to the correct original span.