Open GraphReader

This session presents a practical and production-focused implementation of a GraphRAG architecture that transforms long, complex documents into an explorable knowledge graph and enables intelligent multi-hop retrieval using agentic workflows. We introduce GraphReader, a graph-based retrieval system that improves RAG accuracy by structuring documents into interconnected Section, Fact, and Entity nodes stored in Neo4j. The talk covers an end-to-end ingestion pipeline capable of handling structured, semi-structured, and unstructured documents through dynamic chunking, hierarchical standardization, LLM-driven enrichment, and idempotent document processing.

On top of this graph foundation, we demonstrate a LangGraph-orchestrated retrieval agent that performs hybrid semantic + keyword initial search, omni-directional graph exploration, relevance pruning, and adaptive query reformulation. This multi-hop agentic approach allows the system to capture deeper contextual relationships, significantly reduce hallucinations, and generate grounded, explainable answers.

We discuss implementation details, performance optimizations, and real-world use cases ranging from complex enterprise Q&A and investigative research to scientific literature analysis. The session concludes with key benefits, lessons learned, and future enhancements such as richer relationship extraction, temporal reasoning, and graph-driven validation. Attendees will leave with a clear blueprint for building robust GraphRAG pipelines that enable more accurate, transparent, and context-aware retrieval in modern AI systems