raphaelmansuy/edgequake

Implements LightRAG's knowledge graph approach by decomposing documents into entities and relationships extracted via LLM, then traverses both vector embeddings and graph structures at query time for multi-hop reasoning. Offers 6 query modes (naive vector search through graph-traversing hybrid queries), production-ready PDF processing with optional vision-LLM support for complex layouts, and SQL-level metadata pre-filtering with GIN indexes to eliminate wasted vector scans. Built on Tokio async runtime with PostgreSQL/pgvector backend, OpenAPI REST API, and React 19 frontend with interactive graph visualization.

1,539 stars. Actively maintained with 62 commits in the last 30 days.