A legal research AI that reasons over relationships, not just keywords

The brief was to build a retrieval system for legal research that goes beyond flat keyword search, capable of understanding how legal documents and concepts relate to each other.

The system combines a FastAPI backend, a vector database for semantic search, and a graph database for mapping relationships between documents, all orchestrated through Claude. Hundreds of source documents were ingested and cleaned, including stripping filename issues that broke downstream processing, with token and sentence level chunking strategies tested against each other for retrieval quality. The graph layer is populated and ready for relationship-aware queries on top of the existing semantic search.

A research tool moving from search that finds documents to search that understands how documents relate to each other, built and iterated in stages rather than promised all at once.