AI Agent-Driven Semantic Parsing: Translating Natural Language to Structured Filters with LLMs in E-Commerce

Abstract

Intelligent algorithms that can understand user intent and create structured, dynamic filters to improve product discovery are becoming more and more important for retail e-commerce platforms. Translating unstructured natural language queries into structured attribute-value filters, including category, price, and colour, is the goal of this paper's AI agent-driven architecture that is powered by large language models (LLMs). A user interface layer, a k-nearest neighbour (kNN) catalogue retriever, an LLM-based semantic agent, and a structured filter generator are all part of the system architecture that allows for explainable search interactions in real-time. The approach accomplishes grounded and context-aware parsing by using retrieval-augmented generation (RAG), catalog-aware semantic embeddings, and chain-of-thought prompting. Our technique outperforms standard baselines on a simulated dataset with 50,000 fashion-related queries and 10,000 catalogue items. No-result rates are reduced by 47%, attribute coverage is increased by 21%, and filter accuracy is improved by 22%. In commercial e-commerce settings, our study establishes the groundwork for conversational search systems that are scalable, multimodal, and personalised.