Anthropic Contextual Retrieval targets RAG’s lost-context problem
A Juejin AI article relays Anthropic Engineering’s Contextual Retrieval method, which adds document-level explanations to chunks after splitting and before embedding or keyword indexing. Anthropic reports that Contextual Embeddings combined with Contextual BM25 reduced retrieval failures by 49%, while adding reranking reduced failures by up to 67% in its tests. These benchmark figures are vendor claims relayed by a secondary source.
The method directly addresses a common RAG weakness: chunks can lose the document context needed to retrieve them reliably.
你的 AI 知识库为什么答非所问?问题出在“切块”这一刀
你有没有遇到过这种情况?
公司花了大力气搭知识库,接上大模型。结果一问具体问题,AI 要么答非所问,要么说“没找到”。
文档明明就在库里。
#Anthropic#Contextual Retrieval#Contextual Embeddings#Contextual BM25#Claude 3 Haiku#Gemini#Voyage