Model Context Protocol for Dynamic Shopping Agents: Eliminating Hallucinations in E-Commerce RAG Pipelines

Abstract

Context misalignment remains a critical failure mode in retail chatbots, where discrepancies between user intent, retrieved data, and policy constraints lead to unreliable responses. This paper introduces the Model Context Protocol (MCP), a JSON-schema wrapper that cryptographically binds user queries, retrieval context, and policy rules into a tamper-proof payload. Evaluated on a 75,000-SKU Amazon dataset and 30,000 synthetic queries, MCP reduces hallucinated SKU IDs by 92% and improves first-try resolution rates from 71% to 89%. Technical innovations include the A2A Orchestrator for policy enforcement (38% reduction in fraud false-positives) and a FAISS-based vector store enabling sub-100ms semantic search. We argue that MCP’s modular architecture, open-source validation tools, and cryptographic integrity checks position it as a mandatory layer for production-grade shopping assistants.