Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/avnlp/agentic-med-diag/llms.txt

Use this file to discover all available pages before exploring further.

Agentic Graph RAG for Medical Diagnosis is an open-source system that answers complex clinical questions by combining semantic retrieval and knowledge graph traversal in parallel. Built on LangGraph, it decomposes questions into sub-queries, runs two iterative retrieval channels concurrently, and synthesizes a final evidence-backed answer — all driven by a stateful agentic loop.

Introduction

Understand what the system does, why it was built, and how the pieces fit together.

Quickstart

Set up the environment, configure backends, and run your first medical query.

Architecture

Explore the LangGraph state machine, parent/child subgraphs, and dual retrieval channels.

Backends

Compare LightRAG, MiniRAG, PathRAG, and HyperGraphRAG — switchable at runtime.

How It Works

The system is implemented as a hierarchical LangGraph state machine. A parent graph fans out to two parallel subgraphs — the semantic channel and the relational channel — then synthesizes their outputs into a final answer.
1

Install dependencies

Install uv and run uv sync to install all Python dependencies, including LangGraph, Neo4j, Milvus, and your chosen Graph RAG backend.
2

Configure storage

Start Neo4j for knowledge graph storage and Milvus for dense vector embeddings. Set connection credentials in your environment.
3

Choose a backend

Select from LightRAG, MiniRAG, PathRAG, or HyperGraphRAG. Each backend offers a different graph construction and retrieval strategy, all switchable at runtime.
4

Run a query

Pass a clinical question to the agentic pipeline. The system decomposes it, runs parallel retrieval across both channels, and returns a synthesized, evidence-backed answer.

Key Features

Agentic Multi-Hop Reasoning

Iterative sub-query decomposition with back-references (#N) across hops enables complex multi-step clinical reasoning.

Dual Retrieval Channels

Parallel semantic (text chunks) and relational (SPO triples) channels are run simultaneously and merged before synthesis.

Four Graph RAG Backends

LightRAG, MiniRAG, PathRAG, and HyperGraphRAG — each with distinct graph construction, indexing, and retrieval strategies.

Medical Benchmarks

Evaluated on HealthBench, MedCaseReasoning, MetaMedQA, and PubMedQA using DeepEval metrics.

Build docs developers (and LLMs) love