Resource-efficient retrieval-augmented question answering for the Indian Lok Sabha dataset

The Indian Lok Sabha generates a continuously expanding corpus of legislative records, predominantly archived as unstructured PDF files. Effective public access remains limited due to the shortcomings of keyword-based retrieval systems and the hallucination risks of general-purpose Large Language Models (LLMs).

This paper presents a domain-specific, resource-efficient Retrieval-Augmented Generation (RAG) framework employing DistilGPT-2 (82M parameters) as the generative model, grounded via FAISS-based semantic retrieval using Sentence-BERT embeddings. The pipeline integrates multi-stage PDF preprocessing, semantic indexing, and context-aware response generation. Evaluation was conducted on 450 queries spanning simple, complex, and compound categories, assessed by human annotators using factual accuracy and a five-point relevance scale.

The proposed RAG + DistilGPT-2 framework achieves 94% factual accuracy and a relevance score of 4.6 out of 5, substantially outperforming zero-shot baselines (80% factual accuracy without RAG), while maintaining an average end-to-end inference latency of 1,800 milliseconds (ms) on standard CPU hardware.

The results demonstrate that combining domain-specific retrieval with a lightweight generative model effectively mitigates hallucination and reduces computational overhead, offering a scalable, transparent solution for e-governance applications without reliance on GPU infrastructure.