World's Fair 2025
Practical GraphRAG: Making LLMs smarter with Knowledge Graphs — Michael, Jesus, and Stephen, Neo4j
Overview
This talk introduces Graph RAG, a method for enhancing Large Language Models (LLMs) by integrating knowledge graphs. It addresses the limitations of standard LLMs, such as a lack of domain-specific knowledge, tendency to hallucinate, and difficulty in explaining answers. Graph RAG aims to provide more accurate, contextual, and explainable responses by leveraging structured data within knowledge graphs, moving beyond the limitations of purely vector-based retrieval.
Who should watch
- AI engineers and builders looking to improve LLM accuracy and explainability.
- Product Managers seeking to implement more robust AI solutions.
- Developers working with large datasets and complex domain knowledge.
- Anyone interested in overcoming LLM hallucinations and data bias.
Key takeaways
- Standard RAG systems using vector databases can fall short due to reliance on vector similarity, which doesn't always equate to relevance, and they may only retrieve a fraction of the necessary information.
- Knowledge graphs, composed of nodes, relationships, and properties, offer a structured way to represent rich, complex data, enabling more relevant and contextual retrieval than simple vector similarity.
- Graph RAG combines the logical reasoning capabilities of knowledge graphs with the creative extrapolation of LLMs, analogous to the human brain's left and right hemispheres.
- Constructing a knowledge graph involves multiple steps: substructuring unstructured information into a lexical graph, extracting entities and relationships using LLMs, and enriching the graph with algorithms.
- The retrieval phase in Graph RAG involves an initial index search to find entry points, followed by traversing relationships to fetch additional context, which can be tailored based on user context.
- Research and early studies indicate that Graph RAG can lead to significant improvements in accuracy and potentially reduce token costs compared to traditional RAG methods.
- Tools and open-source projects are available to facilitate knowledge graph construction from various sources like PDFs and web articles, and to implement Graph RAG retrieval pipelines.
- Agentic approaches can break down user questions into tasks, utilizing domain-specific retrievers and tools in a loop to gather and process information for richer LLM responses.
Notable quotes
*Vector similarity is not the same as relevance.*
*We want to do better than this with graph rag and figure out how we can use domain specific knowledge accurate contextual and explainable answers.*
*Knowledge graphs and AI are solving these problems.*
Unofficial community note. Prefer the recording for nuance.