Retrieval works better when the chunk has some context

Anthropic described Contextual Retrieval: adding a short, chunk-specific explanation before embedding a passage and building its BM25 index.

The problem is familiar. A paragraph extracted from a long document may lose the company name, time period, or subject needed to interpret it. Adding that context can make retrieval less ambiguous.

Demo takeaway: evaluate what your search returns before optimizing how your model writes.