Abstract

Colleges in India enrol students from diverse linguistic backgrounds who frequently blend regional languages with English in their communications, a practice commonly termed code-mixing. When students inquire about payment deadlines, such as “mera exam fee kab bharna hai?” (Hinglish) or “naa exam fee enti, ela pay cheyali?” (Telugish), the vast majority of deployed academic chatbots fail, being English-only, context-blind and unprepared. OneVoice is a multilingual chatbot, powered by AI, currently in production at Stanley College of Engineering and Technology for Women, Hyderabad. The system uses a Retrieval-Augmented Generation (RAG) architecture that extracts relevant content from a MongoDB Atlas knowledge base, passed on to the LLaMA 3.3-70B large language model using the Groq inference API. Language and code-mix detection works at the Unicode character level and requires no external NLP library. OneVoice offers five Indian languages and four code-mixed variations including Hinglish, Urdulish, Telugish and Tamlish. Multi-turn dialogue stays coherent for up to 10 turns via context memory. The administrator dashboard, with six analytical tabs, manages sessions, language analysis, knowledge base PDF upload, student feedback and CSV export. Across 20 structured test cases, an overall accuracy of 90% was observed, rising to 100% when ignoring infrastructure rate-limit constraints. The complete system is deployed at zero infrastructure cost using Vercel, Render and MongoDB Atlas free tiers, making it directly replicable by low-resource academic institutions.

Authors

T. Harika, Kaniz-E-Fatima, Sofia Samreen
Stanley College of Engineering and Technology for Women, India

Keywords

Multilingual Chatbot, Code-Mixing, Retrieval-Augmented Generation, Large Language Model, LLaMA