Source: Anthropic cookbook: Contextual Retrieval guide

Runnable Anthropic cookbook that reproduces contextual retrieval on 9 codebases, with the contextualizer prompt, prompt-caching cost stats and Pass@k numbers per stage.

Contents
  • What: vendor notebook / tutorial (Claude cookbook)
  • Author/publisher: Anthropic
  • Date: read 2026-11-14 (recorded run used an earlier Haiku model than the one now named in the code)
  • Original: Enhancing RAG with contextual retrieval

Summary

A hands-on companion to the contextual retrieval blog post. It builds a baseline RAG over 737 chunks from 9 codebases, evaluates it on 248 queries that each have a golden chunk, then adds contextual embeddings, contextual BM25 (Elasticsearch) and a Cohere reranker, reporting Pass@k at each step. It also logs the prompt-caching savings of generating the per-chunk context.

Key claims

  • Metric: Pass@k, whether the golden chunk is in the top k.
  • Pass@10: baseline 87.15%, contextual embeddings 92.34%, reranked pipeline 95.26%. Pass@20 goes from 90.06% (baseline) to 97.45% (full pipeline).
  • Contextualizer prompt: whole document in a <document> block (marked with cache_control), the chunk in a <chunk> block, and a request for a short, succinct situating context, returned with nothing else. Example output: about 76 tokens from about 3.4k input tokens.
  • Process chunks document by document so the cached document prefix (5-minute cache) is reused; in the recorded run 61.83% of input tokens were cache reads, cutting estimated cost from about $9.20 to $2.85 (about 69%).
  • Hybrid fusion: weighted reciprocal rank fusion over the top 150 of each retriever, 0.8 semantic / 0.2 BM25 by default; both raw and contextualized text indexed.
  • Reranking: over-retrieve 10x, rerank with Cohere rerank-english-v3.0; about 100-200 ms and about $0.002 per query added.
  • Caveat: contextualized chunks can exceed an embedding model’s input limit and be truncated.
  • Recommendation tiers: contextual embeddings only for cost-sensitive high volume; add reranking for maximum accuracy.

Feeds

Contextual retrieval

Suggest changes →