Compare chunk overlap with a retrieval benchmark

Overlap can preserve evidence across a chunk boundary, but it also creates more vectors, duplicate retrievals and cost. Choose it from the task rather than copying a default.

Run a boundary case

The RAG ingestion lab creates 11 tokens, five-token chunks and two lexical queries.

from ingestion_cases import overlap_case

result = overlap_case()
assert result["benchmark_queries"] == 2
assert result["scores"]["0"]["queries_hit"] == 1
assert result["scores"]["2"]["queries_hit"] == 2
assert result["selected_overlap"] == 2

Two-token overlap keeps approval exception together across the original boundary and hits both questions. It produces four chunks versus three without overlap. This deterministic lexical fixture does not measure a live embedding retriever.

Compare more than recall

For a realistic corpus, freeze labelled question/evidence pairs and test zero, small and larger overlaps with the same chunk size and retriever. Measure recall@*k*, reciprocal rank, number of chunks, duplicate content in top results, index bytes, ingestion time and downstream answer support.

Inspect by question type. Cross-paragraph questions may benefit; table lookups or self-contained sections may not. If a larger overlap only repeats boilerplate, it can crowd useful evidence out of the retrieval budget.

Keep selection reproducible

Record tokenizer, chunker version, embedding/retriever version and tie-break rule. The lab selects maximum questions hit; a production decision might require a minimum recall then choose fewer chunks or lower cost. Preserve per-case transitions and rerun after changing document structure or model.

Overlap is not a substitute for structural parsing. Keep headings, list items and table rows intact first. Use parent/child retrieval or neighbouring-span expansion when the task needs broader context without indexing every repeated token.

The Generative & Agentic AI course connects overlap to evidence coverage and cost as part of the RAG evaluation report.

Exercise

Create 40 labelled questions and compare three overlap values. Plot recall against index size, inspect duplicate top results and predeclare a tie-break. Retain the failing questions that justify the selected value.

Continue learning

This article is part of the RAG ingestion and document preparation sequence. Use the neighbouring tasks when you need the prerequisite or the next application.

• Review the prerequisite or neighbouring task in Chunk documents around meaningful boundaries.
• Continue with Preserve tables during document ingestion.

Reference: Qdrant points and payload documentation.