GraphReader Agentic RAG: Rethinking Long-Context Retrieval Systems

Session Track: AI Engineering

Session Time: 06:30 am - 07:00 am November 6

Session description

This session focuses on the implementation of GraphReader, a graph-based retrieval system designed to enhance RAG accuracy and performance by structuring long documents into explorable knowledge graphs. The talk will focus on: - GraphReader: How it aims to enable AI agents to retrieve structured information from a document-structured knowledge graph, optimising retrieval with an agentic approach and improving answers to complex queries. - Knowledge Graph Structure: The system breaks documents into smaller text chunks, extracting atomic facts and linking them to key concepts for better information retrieval. - Implementation Details: The GraphReader is built using Neo4j (for graph storage) and LangChain/LangGraph (for defining agent workflows). Code snippets for setting up the Neo4j database, extracting atomic facts, and constructing the knowledge graph. - Agent Exploration Process: The GraphReader agent follows a structured workflow to traverse the graph, starting with key elements, gathering atomic facts, reading text chunks, and optimising retrieval with an agentic approach before providing answers. - Performance Optimisation: The implementation leverages constraints for faster retrieval and indexing, along with a structured prompting approach to guide the AI in selecting relevant facts and key elements.