Agentic GraphRAG: Autonomous Knowledge Graph Construction and Adaptive Retrieval
summary
A multi-agent pipeline automatically infers schema and builds a Neo4j knowledge graph (using spaCy, a BAAI sentence transformer and lightweight LLM components) with seven specialized agents for schema inference, entity/relation extraction, alignment, conflict resolution and confidence scoring. Tests on HotpotQA-scale data (tens of documents, ~25–100) compare FAISS-backed vector retrieval, graph traversal and a hybrid retrieval (vector first then graph expand); hybrid wins, achieving Exact Match 68%. Key failures: missing intermediate entities break graph hops and LLM answering hallucination sets an upper ceiling.
AI-written summary · about
chapters (20)
- 0:00Introduction to Agentic GraphRAG
- 1:04Traditional RAG vs GraphRAG
- 2:31Challenges with GraphRAG
- 3:34Agentic Knowledge Graph Construction
- 5:01Agent Pipeline Breakdown
- 7:30Retrieval Methods Comparison
- 8:33Experiment Setup and Results
- 10:21Key Findings and Conclusions
- 11:46Q&A Session Introduction
- 12:22Handling Supplementary Documents
- 13:38Data Chunking Strategy
- 14:12Sentence Transformer Used
- 15:05Schema Evaluation Method
- 15:27Improving Knowledge Graph Quality
- 16:08Nodes and Edges Count
- 16:51Dataset Source HotpotQA
- 17:28Trusting Results and Metrics
- 18:12Agent Chat History Storage
- 18:55Training the Retrieval Router
- 20:06Closing Remarks and Code Sharing
related talks
20mEnhancing RAG with Multi-Agent Integration
29mAgentic GraphRAG: Multi-Agent Knowledge Graph Construction for Research Teams
28mAgentic GraphRAG: Multi-Agent Knowledge Graph
36mAgentic GraphRAG: AI’s Logical Edge
27mAdaptive Knowledge Graphs for Dynamic RAG Systems
01h 26mThinking in Graphs: The Cognitive Foundations of Agentic AI
39mEnhancing RAG with Neo4j Knowledge Graph
30mAgentic RAG with Knowledge Graphs Explained