LLM and MCP driven e-commerce assistant over graph databases

Abstract

Traditional e-commerce search-and-filter interfaces often lead to high cart abandonment rates due to rigid keyword matching and complex navigation. This paper presents an architecture for conversational e-commerce that integrates large language models (LLMs), Model Context Protocol (MCP) servers, and a Neo4j graph database to enable natural language product discovery. The system employs an LLM-powered assistant that securely executes Cypher queries over a graph-structured product catalog, with MCP servers enforcing validation to prevent direct database exposure. A near-real-time synchronization pipeline maintains consistency between the Neo4j knowledge graph and the transactional PostgreSQL database. The architecture supports multi-turn conversations with persistent context, enabling personalized recommendations grounded in explicit product relationships. By combining secure orchestration, graph-based reasoning, and conversational context, the proposed system reduces search effort while maintaining transparency and explainability.