Soba Labs guide says fix retrieval before the model in enterprise RAG on LangChain

Kacper Klusek of Soba Labs recommends fixing retrieval before changing the model when building enterprise RAG on LangChain, according to a LangChain post on X dated October 31, 2026.

He points to a retrieval chain combining contextual embeddings, BM25 and a reranker. He cites Anthropic's published evaluation, which credits that combination with cutting failed retrievals by 67%. Klusek also presents LangSmith online evals as a way to catch quality decay when documents change.

The post reports Klusek's recommendations and the Anthropic figure as stated in the post. It does not include independent testing of the approach.