voyage-law-2

voyage-law-2 is a domain-specialized embedding model from Voyage AI built for legal text retrieval. It is documented as leading the MTEB legal retrieval leaderboard, outperforming OpenAI v3 large by roughly six percent on average across eight legal retrieval datasets, with even larger margins above ten percent on LeCaRDv2, LegalQuAD, and GerDaLIR. The model was trained on extensive long-context legal corpora, allowing it to retrieve relevant passages across lengthy documents such as court opinions, contracts, and statutes. Beyond its legal focus, it is reported to match or exceed performance on general-purpose corpora across other domains, making it a reasonable choice when teams want strong legal accuracy without sacrificing broader retrieval quality. The model supports a 16K token context window, which lets it encode substantial sections of a legal filing in a single embedding rather than fragmenting them into small chunks. It is accessible through embedding APIs and can be wired into hybrid search systems alongside lexical search. Practitioners building RAG pipelines for case law, regulatory archives, or compliance knowledge bases benefit most from its specialization, while teams working primarily on short, general text may find newer general-purpose Voyage models more cost-effective.

voyage-law-2 is a domain-specialized embedding model from Voyage AI built for legal text retrieval. It is documented as leading the MTEB legal retrieval leaderboard, outperforming OpenAI v3 large by roughly six percent on average across eight legal retrieval datasets, with larger margins above ten percent on LeCaRDv2, LegalQuAD, and GerDaLIR. Training on extensive long-context legal corpora allows it to retrieve relevant passages across lengthy documents such as court opinions, contracts, and statutes. Beyond its legal focus, it is reported to match or exceed performance on general-purpose corpora across other domains, giving teams strong legal accuracy without sacrificing broader retrieval quality. The model supports a 16K token context window, enabling it to encode substantial sections of a legal filing in a single embedding rather than fragmenting them into small chunks. It is accessible through embedding endpoints and can be wired into hybrid search pipelines alongside lexical search. Practitioners building retrieval-augmented systems for case law, regulatory archives, or compliance knowledge bases benefit most from its specialization, while teams working primarily on short, general text may find newer general-purpose Voyage models more cost-effective.