DocGraph

Turn documents into an interactive knowledge graph you can talk to.

DocGraph ingests documents, extracts entities and relationships, renders them as an interactive force-directed graph, and answers questions over the corpus using an Agentic RAG pipeline that returns answers with citations.

Architecture

This is a monorepo with two deployables:

Package	Stack	Responsibility
docgraph-web	Next.js / TypeScript, react-force-graph	Upload UI, interactive knowledge-graph visualization, chat
docgraph-agent	TypeScript, Flue agent, Gemini, PostgreSQL + pgvector	Entity/relationship extraction, embeddings, Agentic RAG Q&A with citations

Highlights

• Agentic RAG with citations — answers are grounded in retrieved passages, not free-form generation.
• Knowledge graph extraction — documents become entities + relationships, rendered with react-force-graph.
• Flue agent — orchestrates extraction and retrieval as a programmable agent harness.
• Vector search — PostgreSQL + pgvector for semantic retrieval.
• Multi-tenant — OAuth auth, per-tenant isolation.
• Deployed on Fly.io.

Getting started

Each package is self-contained — see docgraph-web/ and docgraph-agent/ for setup, environment variables, and run instructions.