Chunk overlap means adjacent chunks share a small amount of duplicated text at their boundary — a simple technique to reduce the risk of losing meaning that spans across a chunk split.
Why Overlap Helps
Without overlap:
Chunk 1: "...the refund policy allows returns within 30 days"
Chunk 2: "of purchase, provided the item is unused. Sale items..."
→ A query about "how many days for a refund" might only
retrieve Chunk 1, missing the "of purchase" continuation,
or might only retrieve Chunk 2, missing the "30 days" part
entirely — the key fact is split across the boundary.
With overlap (e.g. 50 tokens shared):
Chunk 1: "...the refund policy allows returns within 30 days
of purchase, provided the item is unused."
Chunk 2: "of purchase, provided the item is unused. Sale
items are final sale and cannot be returned."
→ Either chunk retrieved alone still contains the complete
"30 days of purchase" fact.How Much Overlap Is Reasonable?
A common starting range is roughly 10-20% of the chunk size, though — as with chunk size itself — this is a parameter worth tuning empirically for your specific content rather than treating as a fixed rule. Too little overlap risks the boundary problem above; too much overlap wastes storage and retrieval budget on duplicated content.
Overlap Isn't a Complete Fix
Overlap reduces, but doesn't eliminate, the boundary-splitting problem — a fact that spans a distance longer than the overlap window can still get split. Structure-aware chunking (splitting at natural boundaries like paragraph or section breaks, see Chunking Strategies) addresses the root cause more directly than overlap alone.
Practical Use Case
Content with important facts that could plausibly span typical chunk boundaries (numbered lists, multi-sentence policy statements, step-by-step instructions) benefits meaningfully from overlap — content that's already naturally broken into short, self-contained pieces (like FAQ entries) needs less.
Common Mistakes
- Using zero overlap by default without considering the boundary-splitting risk for your specific content
- Using excessive overlap "to be safe," inflating storage and retrieval costs without a measured quality benefit
- Relying on overlap alone instead of also considering structure-aware chunking for content where boundaries genuinely matter
Interview Relevance
"What problem does chunk overlap solve, and does it fully solve it?" — reducing (not eliminating) the risk of splitting important content across a chunk boundary is the accurate, complete answer.
Practice Question
A user asks "what's step 3 of the setup process" and gets an incomplete answer because step 3's instructions were split across two chunks. Propose a fix involving overlap, and explain its limitation.