Measured, then written down

4 October 2026·Research·9 min read

Is Contextual Retrieval worth it? We tested it against BM25, similarity search, reranking and filters

We tested Anthropic's Contextual Retrieval on a deliberately hard archive of 287 documents, side by side with keyword search, BM25, similarity search, a reranker and metadata filters. Here is what each one added to recall, and what happened when we used them all together.