What is chunk overlap? Why adjacent passages repeat and what it costs
Understand overlap through cross-boundary qualifications, including target versus actual overlap, continuity, duplicate retrieval, and context usage.
Understand the concept
Chunk overlap repeats content between adjacent passages to reduce boundary losses. It does not guarantee complete context, and more is not always better. Repetition increases indexed content and duplicate retrieval, while actual overlap depends on splitter rules. Inspect produced chunks rather than configuration alone.
Conditions near a boundary may need repetition
A scope condition at one paragraph’s end can qualify instructions in the next. Overlap can preserve continuity, but distant prerequisites may still need headings or structure rather than unlimited repetition.
Microsoft discusses overlap by content type, and LangChain describes chunk_overlap as a target. Paragraph-preserving rules can produce actual overlap different from the setting. This explanation uses no real corpus and prescribes no universal ratio.
Duplicate hits can crowd out other evidence
Repeated sentences can fill top results with neighboring chunks while excluding other exceptions. More chunks do not imply broader evidence. Inspect source positions and check whether deduplication preserves context.
Overlap adds embedding and storage work and can consume context when concatenated repeatedly. One improved example does not justify a corpus-wide increase. Tables and short independent notes may need different treatment.
Choose settings by inspecting chunks and fixed questions
Compare no overlap and candidate settings on boundary-crossing questions, then add ordinary questions to check noise. This is a test design without measured results. Record units and separators to avoid comparing characters with tokens.
Duplicating text cannot restore headers lost during parsing. Check full answer qualifications rather than chunk counts alone. Overlap is one continuity mechanism to evaluate alongside chunking and retrieval.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.