RManLuo/gfm-rag

Combines graph neural networks with knowledge graph indexing to enable multi-hop reasoning within a single retrieval step, eliminating the need for fine-tuning on new datasets. Constructs a knowledge graph index from documents to capture entity relationships, then uses a pre-trained GFM retriever to reason over the graph structure and surface relevant documents for LLM generation. Integrates with arbitrary agent frameworks and offers interpretable retrieval paths while supporting domain-specific fine-tuning for specialized tasks.

222 stars.