An implementation of multi-hop, agentic Graph Retrieval-Augmented Generation (Graph RAG) for medical question answering. The system combines semantic and relational retrieval channels - backed by Neo4j and Milvus - to answer complex clinical questions that require multi-hop reasoning across a medical knowledge graph.

Code coming soon.

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.

Full agentic multi-hop pipeline:

• Retrieves initial context via a GraphRAG backend
• Decomposes the question into semantic sub-queries (with #N back-references) and relational SPO triple queries (with Entity#N placeholders)
• Runs two parallel iterative retrieval-and-reasoning loops:
  • Semantic channel: sub-query grounding → GraphRAG retrieval → semantic filter → text summary → sub-answer → logic drafting → evidence verification → conditional expansion
  • Relational channel: SPO triple queries → KG filter → triplet list summaries → sub-answer
• Synthesizes a final answer combining both channels

• Neo4j — stores the entity/relationship knowledge graph built from medical documents
• Milvus — stores dense vector embeddings for semantic retrieval over document chunks

Paper: https://arxiv.org/abs/2410.05779
GitHub: https://github.com/HKUDS/LightRAG

LightRAG builds a dual-level knowledge graph — a local graph of fine-grained entity mentions and a global graph of high-level concept clusters — and retrieves context from both simultaneously in hybrid mode.

Indexing:

• Documents are chunked and passed to an LLM to extract named entities and their pairwise relationships
• Entities and relationships are stored in Neo4j, with embeddings stored in a vector index alongside raw document chunks
• Parallel async inserts and async LLM calls accelerate large-scale ingestion

Retrieval (hybrid mode):

• Performs both keyword-based KG traversal and dense vector search over document chunks
• Returns a structured JSON context with four sections: Knowledge Graph Data (Entity), Knowledge Graph Data (Relationship), Document Chunks, and a Reference Document List
• The context_filter splits this into the semantic channel (document chunks + reference list) and the relational channel (entity + relationship JSON)

Paper: https://arxiv.org/abs/2501.06713
GitHub: https://github.com/HKUDS/MiniRAG

MiniRAG is a lightweight Graph RAG designed for resource-constrained settings. It uses a simpler graph construction process and a reduced "light" retrieval mode that prioritises efficiency over exhaustive graph traversal.

Indexing:

• Same chunking parameters as LightRAG (400 tokens, 50-token overlap)
• Extracts entities and relationships into a flat graph stored in the working directory (no Neo4j dependency)
• Lower memory and LLM-call overhead than heavier backends

Retrieval (light mode):

• Focuses on the most immediately relevant neighbourhood of the query entity rather than full hybrid traversal
• Returns a CSV-formatted context with three sections: Entities, Relationships, and Sources
• The context_filter splits this into the semantic channel (Sources CSV) and the relational channel (Entities + Relationships CSVs)

Paper: https://arxiv.org/abs/2502.14902
GitHub: https://github.com/BUPT-GAMMA/PathRAG

PathRAG organises the knowledge graph into a two-tier hierarchy — a global layer of abstract, high-level entities and relationships, and a local layer of fine-grained, low-level entities and relationships — and retrieves reasoning paths between entities rather than isolated nodes or edges.

Indexing:

• Constructs both high-level (thematic/concept) and low-level (mention-level) entity and relationship layers during ingestion
• Stores the hierarchical graph in the working directory

Retrieval (hybrid mode):

• Traverses connecting paths between query-relevant entities, returning chains of evidence rather than individual triples
• Returns a CSV context with five sections: high-level entity information, high-level relationship information, Sources, low-level entity information, and low-level relationship information
• The context_filter separates the semantic channel (Sources CSV) from the relational channel (both high-level and low-level entity/relationship CSVs)

Paper: https://arxiv.org/abs/2503.21322
GitHub: https://github.com/LHRLAB/HyperGraphRAG

HyperGraphRAG extends the standard pairwise-edge graph model with hyperedges — edges that connect more than two entities at once. This allows a single relationship to capture group interactions among multiple medical concepts simultaneously (e.g., a clinical syndrome that jointly implicates multiple symptoms, biomarkers, and treatments).

Indexing:

• Extracts entities and higher-order relationships (hyperedges) from documents during ingestion
• Represents multi-entity relationships as hyperedges in the graph, stored in the working directory

Retrieval (hybrid mode):

• Combines dense vector search with hypergraph traversal to surface both document chunks and multi-entity relational context
• Returns a CSV context with three sections: Entities, Relationships (including hyperedges), and Sources
• The context_filter is structurally identical to MiniRAG's split — Sources for the semantic channel, Entities + Relationships for the relational channel

We evaluate all four pipelines against LightRAG, MiniRAG, PathRAG, and HyperGraphRAG baselines using DeepEval metrics.

Metrics: Contextual Recall, Contextual Precision, Contextual Relevancy, Answer Relevancy, Faithfulness

Datasets:

This project uses uv for dependency management.

# Install uv
pip install uv

# Install dependencies
uv sync

This project is licensed under the MIT License - see the LICENSE file for details.

• GraphSearch
• LangGraph
• Neo4j
• Milvus
• DeepEval
• LightRAG - GitHub
• MiniRAG - GitHub
• PathRAG - GitHub
• HyperGraphRAG - GitHub