Agentic GraphRAG
A technical architecture that unifies the RAG context layer and the agent memory layer into a single graph structure, comprising a knowledge layer, a working memory layer and a decision memory layer, enabling cross-layer traceability and propagation of decision invalidation.
Timeline (last 90 days)
Agentic GraphRAG versions decision revisions, keeping old versions in history rather than overwriting and deleting them.
Agentic GraphRAG's entity extraction is driven by a user-defined schema, with local extraction using GLiNER and structured LLM extraction handled by BAML on selected chunks.
The author acknowledges that Agentic GraphRAG's citation check cannot determine whether the evidence actually supports the conclusion, and semantic correctness still requires human or upstream process review.
Agentic GraphRAG provides an MCP server, so any Agent runtime that supports MCP can call its graph retrieval and memory capabilities without being tied to DeepAgents/LangGraph.
If the scenario is just simple question answering, the cost of introducing a three-layer graph and schema modeling may not be worthwhile; the three-layer structure is better suited to scenarios that need to audit an Agent's decision chain (such as compliance review, research reproduction, and long multi-session tasks).
Agentic GraphRAG finds paragraphs through Neo4j's native full-text index and vector index, and uses reciprocal rank fusion (RRF) to fuse keyword rankings and semantic rankings.
Retrieval in Agentic GraphRAG supports "reading beyond the search hit", meaning it can navigate to the section containing the hit paragraph, adjacent paragraphs, or table rows with headers, and entity relationships can point back to the supporting paragraphs.
The Agentic GraphRAG project stores the knowledge layer, working memory layer and decision memory layer in the same Neo4j graph, with the three layers connected to each other by edges.
When a referenced document changes, Agentic GraphRAG marks the affected decisions and their explicit dependencies as pending review.
The roadmap of the Agentic GraphRAG project includes: modularizing the pipeline to support different hardware, extending retrieval to more graph databases and storage backends, and running large-scale benchmark tests on domain-specific datasets, with results to be released soon.
5 more timeline events
All Facts (15)
If the scenario is just simple question answering, the cost of introducing a three-layer graph and schema modeling may not be worthwhile; the three-layer structure is better suited to scenarios that need to audit an Agent's decision chain (such as compliance review, research reproduction, and long multi-session tasks).
50%UnverifiedAgentic GraphRAG versions decision revisions, keeping old versions in history rather than overwriting and deleting them.
50%UnverifiedAgentic GraphRAG's entity extraction is driven by a user-defined schema, with local extraction using GLiNER and structured LLM extraction handled by BAML on selected chunks.
50%UnverifiedThe author acknowledges that Agentic GraphRAG's citation check cannot determine whether the evidence actually supports the conclusion, and semantic correctness still requires human or upstream process review.
50%UnverifiedAgentic GraphRAG provides an MCP server, so any Agent runtime that supports MCP can call its graph retrieval and memory capabilities without being tied to DeepAgents/LangGraph.
50%UnverifiedAgentic GraphRAG finds paragraphs through Neo4j's native full-text index and vector index, and uses reciprocal rank fusion (RRF) to fuse keyword rankings and semantic rankings.
50%UnverifiedRetrieval in Agentic GraphRAG supports "reading beyond the search hit", meaning it can navigate to the section containing the hit paragraph, adjacent paragraphs, or table rows with headers, and entity relationships can point back to the supporting paragraphs.
50%UnverifiedThe Agentic GraphRAG project stores the knowledge layer, working memory layer and decision memory layer in the same Neo4j graph, with the three layers connected to each other by edges.
50%UnverifiedWhen a referenced document changes, Agentic GraphRAG marks the affected decisions and their explicit dependencies as pending review.
50%UnverifiedThe built-in Agent of Agentic GraphRAG checks whether the referenced paragraph was actually retrieved in this run and whether it belongs to a document the Agent is authorized to read.
50%UnverifiedThe knowledge layer of Agentic GraphRAG uses Docling for document ingestion and Chonkie for chunking, preserving document section structure, tables and paragraph hierarchy during chunking, and PDFs also record page number positions.
50%UnverifiedThe built-in Agent of Agentic GraphRAG is based on DeepAgents/LangGraph and records messages and tool calls as task events, supporting task progress resumption in subsequent sessions.
50%UnverifiedIn the decision memory layer of Agentic GraphRAG, decisions can be reaffirmed or replaced, and historical records are not lost.
50%UnverifiedEntity extraction in Agentic GraphRAG can use a locally running GLiNER model, or BAML can be used for structured LLM extraction on selected chunks.
50%UnverifiedThe roadmap of the Agentic GraphRAG project includes: modularizing the pipeline to support different hardware, extending retrieval to more graph databases and storage backends, and running large-scale benchmark tests on domain-specific datasets, with results to be released soon.
50%