Agentic Commerce: A Unified Multi-Retrieval Framework for High-Fidelity E-Commerce Chatbots

Abstract

E-commerce chatbots face critical limitations that undermine customer trust: hallucinations from ungrounded responses, poor multi-turn coherence, and an inability to execute real-world actions such as processing refunds or verifying live inventory. Existing retrieval-augmented generation (RAG) and chain-of-thought (CoT) approaches address knowledge grounding and reasoning, respectively, yet remain fundamentally passive—they inform but cannot act. We present a unified agentic framework that integrates RAG for factual grounding, CoT for structured multi-step reasoning, and multi-agent col- laboration for autonomous task execution. The modular architecture encompasses specialized agents for retrieval, reasoning, action generation, and safety enforcement, orchestrated through an LLM-based routing policy. This design enables the system to move beyond answering questions toward completing transactions, coordinating inventory checks, and resolving complex customer inquiries autonomously. Evaluated on a 10K-SKU e-commerce dataset spanning factual, comparative, and multi-turn query types, the framework achieves 96.2% response accuracy and 95.8% grounding reliability. The multi-agent ar- chitecture reduces errors in multi-turn interactions by 18% compared to single-agent baselines. The system operates with a median latency of 3.12 seconds—a deliberate safety-first design choice that pri- oritizes transactional reliability over conversational speed, ensuring business-critical accuracy in high- stakes operations where sub-second responses would compromise correctness.