Legal RAG corpus maps 529K sections

This reads like a real-world proof that legal RAG is still mostly a retrieval-engineering problem, not an embeddings problem. The graph layer is the interesting part: once you have clean citations and cross-references, personalization over structure can surface relevant provisions that semantic search alone misses.

• –BM25 matters here because legal queries often hinge on exact section numbers, defined terms, and statutory phrasing
• –Dense retrieval still adds value for cross-jurisdiction similarity, where wording diverges but legal function is the same
• –Citation graphs plus PageRank are the differentiator because they encode how statutes actually relate, not just how they sound
• –The data work is the hard moat: 50 scraper variants, normalization quirks, and edge resolution are where most teams would stall
• –MCP exposure makes the corpus immediately useful to agents, which is a cleaner distribution story than shipping yet another search UI