Two brothers from Australia have built what appears to be the best legal document search AI currently available. Isaacus, founded by Umar and Abdur-Rahman Butler, just released three models under the Kanon 2 banner, and the benchmark results back up the claim.
The Kanon 2 Embedder and Kanon 2 Reranker work together to find and rank relevant legal documents. Embedders convert text into numerical representations so an AI can compare documents by meaning rather than keywords. Rerankers then take a rough set of results and re-sort them by actual relevance. Together, these two models rank first on both Legal RAG Bench and the Massive Legal Embedding Benchmark (MLEB), which are the primary ways researchers measure how well AI can retrieve legal information.
The third release, Kanon 2 Enricher, is more unusual. It is described as an "entirely new type of AI model" focused on enriching legal data, though Isaacus has not published full details on how it differs from standard approaches.
Legal search is a space where generic AI models consistently underperform. Legal language is dense, citation-heavy, and full of domain-specific meanings that trip up general-purpose tools. That is why vertical-specific models like these matter for lawyers and legal researchers who need to find the right case law or statute quickly, not just plausible-sounding results.
Isaacus is still a small, early-stage company competing against well-funded players like CoCounsel (backed by Thomson Reuters) and Harvey AI. Topping benchmarks is a strong start, but benchmark performance and real-world reliability in legal work are not the same thing. The real test will be whether law firms trust these models enough to build workflows around them.