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

Area of Science:

• Natural Language Processing
• Information Retrieval
• Computational Linguistics

Background:

• Legislative records from the Indian Lok Sabha are primarily unstructured PDFs, limiting public access.
• Existing keyword search and general Large Language Models (LLMs) struggle with accuracy and hallucination for these documents.

Purpose of the Study:

• To develop a domain-specific, resource-efficient Retrieval-Augmented Generation (RAG) framework for accessing legislative records.
• To mitigate hallucination risks and improve factual accuracy in information retrieval from unstructured government documents.

Main Methods:

• A RAG framework was implemented using DistilGPT-2 (82M parameters) and FAISS-based semantic retrieval with Sentence-BERT embeddings.
• The pipeline included multi-stage PDF preprocessing, semantic indexing, and context-aware response generation.
• Evaluated on 450 queries, assessing factual accuracy and relevance via human annotators.

Main Results:

• The RAG + DistilGPT-2 framework achieved 94% factual accuracy and a 4.6/5 relevance score.
• Outperformed zero-shot baselines (80% factual accuracy) significantly.
• Maintained an average inference latency of 1,800 ms on standard CPU hardware.

Conclusions:

• Domain-specific retrieval combined with a lightweight generative model effectively reduces hallucination and computational costs.
• The framework offers a scalable, transparent solution for e-governance, not requiring GPU infrastructure.