Anthropic Introduces Contextual Retrieval with 67% Improvement to Boost RAG Systems

Sep 20, 2024, 05:55 AM

Anthropic has introduced a new technique called Contextual Retrieval, which significantly improves the performance of Retrieval-Augmented Generation (RAG) systems. This method reduces incorrect chunk retrieval rates by up to 67% when combined with reranking and achieves a 49% improvement in retrieval accuracy. The technique involves generating contextualized chunks at a cost of $1.02 per million document tokens, using Contextual Embeddings and Contextual BM25 to enhance retrieval accuracy. Contextual Retrieval, when combined with prompt caching, is expected to transform AI assistants in sectors such as hospitality, healthcare, and finance by making them smarter and more efficient.