MRAG-HC: Multilingual Retrieval-Augmented Generation System with Hallucination Control Mechanism

MRAG-HC is an end-to-end multilingual Retrieval-Augmented Generation (RAG) platform developed by master's students from the Department of Computer Science and Engineering at VNIT Nagpur, India. It supports three languages: English, Hindi, and Marathi, integrates OCR and FAISS vector database, with the core goal of reducing hallucination issues in Large Language Models (LLMs). This project, as a degree thesis achievement, was released in May 2028.

Although LLMs have made progress in the field of natural language processing, they have the "hallucination" problem—generating false or unsubstantiated content, which is extremely harmful in scenarios requiring factual accuracy such as healthcare and law. Retrieval-Augmented Generation (RAG) reduces hallucinations by retrieving context from external knowledge bases, but traditional RAG may still produce hallucinations when the retrieved documents do not fully match the query.

The core innovations of MRAG-HC include native multilingual support (English, Hindi, Marathi), hallucination control mechanism, end-to-end pipeline, OCR integration, and FAISS vector database. Multilingual support is achieved through multilingual embedding models (e.g., multilingual-e5), language detection, and cross-language retrieval, mapping texts in different languages to a unified vector space.

The core RAG process includes document ingestion (supports PDF, images, etc., OCR extracts text from scanned documents), chunking, vectorization, FAISS index construction, query processing, semantic retrieval, reordering, context construction, and generation. Hallucination control uses multi-layer strategies: confidence scoring, source verification (NLI model to judge entailment relations), uncertainty quantification, retrieval-generation alignment, and multi-source cross-validation.

MRAG-HC can be applied to scenarios such as government document query (multilingual policy information), multilingual knowledge base (internal enterprise services), educational assistance (native language academic query), news verification (multilingual content validation), medical information retrieval (literature-based Q&A), etc., and has important social value.

The project faces challenges such as differences in multilingual embedding quality (using Indian language-specific models or fine-tuning general models), accuracy-recall trade-off in hallucination control (balancing conservatism and practicality), OCR error propagation (integrating OCR confidence scoring), computing resource limitations (optimizing inference efficiency), and corresponding solutions are adopted.

Current limitations include insufficient language coverage (only three languages), lack of standardized benchmarks for hallucination evaluation, and no optimization for large-scale production. Future plans include expanding to more Indian languages, supporting multimodal content, real-time knowledge base updates, personalized strategies, and federated learning.

MRAG-HC promotes the development of RAG technology towards multilingualism and high reliability, meets the needs of India's language diversity, ensures information accuracy, and is suitable for high-reliability scenarios. As an academic project, its research insights provide guidance for responsible AI applications, contributing technical implementation and research value.