Team20-Voice-Based Agentic Ecommerce in Regional Indian Languages

Indian Institute of Science Summer 2025 DA225o Submission17 Authors

Keywords: RAG, SarvamAI, LangChain, LangGraph, Agent, Vector Embeddings

TL;DR: Implement LangGraph Agent with RAG for voice first ecommerce

Abstract: This paper presents a novel agentic ecommerce system that addresses the challenges of voice-based shopping in linguistically diverse regions of India. Our solution will enable users to interact with an ecommerce platform using voice messages in regional Indian languages, processing their requests through a specialized pipeline that converts speech to text, identifies product requirements, and facilitates end-to-end purchase experiences. The system employs a Retrieval-Augmented Generation (RAG) architecture specifically designed for ecommerce contexts, which leverages product catalogs to provide contextually relevant responses while maintaining a persistent shopping cart throughout the conversation. Our implementation incorporates three key components: (1) a speech processing system utilizing Sarvam AI's API for handling diverse Indian languages, (2) an intent recognition framework that extracts product details, quantities, and user objectives from conversational inputs, and (3) a context-aware response generation mechanism that maintains conversation coherence across multiple turns. We hope to demonstrate the practical feasibility of this approach through rapid development, detailing the system architecture, component integration, and workflow optimization. Our work contributes to making digital commerce more accessible to linguistically diverse populations with varying levels of technological literacy, providing insights for future research in voice-first commercial interactions in multilingual environments.

Submission Number: 17