What is contextual retrieval, or contextual chunk enrichment?

30-second answerSay your answer out loud first, then reveal.

Chunks often lose meaning when cut out of their document. "Revenue grew 3% over the previous quarter" doesn't say which company or which quarter. Contextual retrieval prepends a short, document-aware context to each chunk before embedding and BM25 indexing, typically generated by an LLM that sees the whole document, so the chunk becomes self-contained. Anthropic's 2024 write-up reported large reductions in retrieval failures, especially when combined with BM25 and reranking.

Example.
Original chunk: "The company's revenue grew by 3% over the previous quarter."
Contextualised: "This chunk is from ACME Corp's Q2 2026 10-Q filing, section on financial results. The company's revenue grew by 3% over the previous quarter."

Implementation

For each chunk, call an LLM with the full document plus the chunk: "Give short context situating this chunk within the document."
Prepend the context to the chunk text.
Embed it and add it to the BM25 index (contextual BM25).
Use prompt caching so the full document is cached across all its chunks, which reduces cost a lot.

Simpler, cheaper variants

Contextual chunk headers: deterministically prepend the document title + section path ("Leave Policy > Casual Leave > Eligibility") without an LLM.
Document summary prepended to each chunk.

Trade-offs: one-time indexing cost per chunk (an LLM call), longer chunks, re-processing when documents change. Usually worth it when documents are long and chunks are context-dependent (financial filings, contracts, manuals).

PreviousExplain parent-child (small-to-big) retrieval and sentence-window retrieval.

You understood something today that you didn't yesterday.

What is contextual retrieval, or contextual chunk enrichment?

Implementation

Simpler, cheaper variants

Related