How to Build Knowledge Graphs Using AI Agents and Vector Search - Demo Overview

Sabika Tasneem discusses a novel approach to resolving redundancy and inconsistency in knowledge graphs by using AI agents and vector search, as demonstrated by Carl Kugblenu during a hackathon at Finland's VTT. Kugblenu developed a pipeline that employs large language models (LLMs) and Memgraph’s vector search to tackle the complex problem of entity disambiguation, ensuring that each unique concept is represented only once in a knowledge graph. The process involves context-aware similarity search analysis, where vector embeddings and cosine similarity are used to identify candidate pairs of mentions, which are then processed by a GPT-powered agent resolution pipeline for merging or node creation. The live demonstration showcased the transformation of a chaotic dataset into a clean, canonical graph, opening up advanced analytics possibilities like PageRank and community detection. Despite challenges such as LLM hallucinations and embedding dimension issues, the project highlighted the importance of metadata-rich contexts and the benefits of combining LLMs with similarity-based approaches, providing valuable insights for organizations aiming to implement similar systems.