Hand-rolled GraphRAG meets DeepSearch in a multi-agent kitchen

What it does This is a kitchen-sink RAG agent stack that rebuilds GraphRAG from scratch and layers private-domain DeepSearch on top of it. Documents in PDF, Markdown, CSV, and other formats get parsed into entities and relationships by LLMs, fed into a knowledge graph with community detection and incremental updates, and queried through a FastAPI backend and web frontend.

The interesting bit The project treats DeepSearch as a graph-native operation rather than a vector-store add-on, adding evidence-chain tracking and reasoning visualization. Its real personality is the Plan-Execute-Report multi-agent committee: a Planner decomposes and reviews tasks, a WorkerCoordinator dispatches retrieval and reflection jobs, and a Reporter assembles long-form answers via outline builders and consistency checks. It also ships a custom evaluation framework with over 20 metrics—unusual depth for a personal integration project.

Key highlights

• From-scratch GraphRAG core plus integrations for LightRAG and Neo4j-llm-graph-builder
• DeepSearch fused with knowledge graphs instead of pure vector retrieval, with multi-path parallel search and evidence tracing
• Plan-Execute-Report agent architecture featuring Clarifier, TaskDecomposer, PlanReviewer, and Map-Reduce report generation
• Incremental graph updates with entity disambiguation, alignment, and conflict resolution
• Debug mode surfaces LangGraph trajectories, retrieved subgraphs, and source documents in the UI

Caveats

• The README describes architecture in depth but delegates the quickstart to a separate start.md, so setup guidance is not visible in the main document
• Specifics on supported LLM providers, hardware requirements, or latency characteristics are absent from the provided sources

Verdict A solid reference if you want to see how graph construction, agentic reasoning, and evaluation fit together in one codebase. Avoid if you are looking for a minimal, drop-in library rather than a full-stack integration demo.