AI-BASED MULTILINGUAL GOVERNMENT SCHEME ASSISTANT

This paper access to government welfare scheme information in India remains inequitable owing to linguistic diversity, fragmented portals, and complex eligibility criteria. This paper presents SchemeBot, an end-to-end multilingual retrieval-augmented generation (RAG) system that enables citizens to query a unified knowledge base of over 1,200 central and state welfare schemes in Hindi, Tamil, Telugu, Kannada, and English through a natural-language conversational interface. The system encodes both scheme documents and user queries using the BAAI/bge-m3 multilingual dense retriever, which maps text into a shared 1024-dimensional semantic space, and retrieves candidate chunks from a FAISS HNSW index optimised for sub-50 ms latency. Retrieved context is passed to a locally hosted Falcon3-3B-Instruct model via a LangChain RAG chain for grounded, citation-aware response generation. We formalise the retrieval objective as a Maximum Inner-Product Search (MIPS) problem and characterise the embedding geometry under cosine similarity. Offline evaluation on a manually curated benchmark of 2,400 query-scheme pairs yields a Retrieval Recall@5 of 0.913, NDCG@5 of 0.871, BERTScore F₁ of 0.847, and end-to-end response latency of 218 ms at the 95th percentile. These results represent statistically significant improvements over TF-IDF and BM25 baselines (p < 0.01, two-tailed Wilcoxon signed-rank test). SchemeBot is deployable on commodity hardware without GPU acceleration, making it suitable for integration into India’s Common Service Centre infrastructure.