Optimizing RAG for Contract Analysis: Our Research Findings
Robin AI optimised its use of Retrieval-Augmented Generation (RAG) for contract analysis by enriching chunks with metadata like clause labels and summaries, boosting retrieval accuracy by up to 6%. Testing showed that only 15% of a contract’s text was needed to reach 90% recall, significantly reducing processing load. Among embedding models, Voyage 3 Large performed best, and pairing optimised RAG with Claude by Anthropic produced answers as accurate as full-document inputs. The result: faster, cheaper, and more precise insights from even the most complex contracts.