Mitigating hallucinations in large language models through enhanced retrieval-augmented generation: A Framework with FAISS and Lang Chain
1 School of Technology Management and Engineering, NMIMS Chandigarh.
2 University Institute of Engineering and Technology, Panjab University, Chandigarh, India.
Research Article
Open Access Research Journal of Engineering and Technology, 2026, 10(02), 058-063.
Article DOI: 10.53022/oarjet.2026.10.2.0033
Publication history:
Received on 20 March 2026; revised on 26 April 2026; accepted on 28 April 2026
Abstract:
Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language understanding and generation, yet they remain susceptible to hallucinations—generating factually incorrect or ungrounded information. This study presents an enhanced Retrieval-Augmented Generation (RAG) framework integrating FAISS vector database, Sentence-BERT embeddings, and Lang Chain orchestration to mitigate hallucinations and improve response accuracy. The proposed system implements a hybrid retrieval pipeline combining dense vector search with semantic reranking. Experimental evaluation on a curated technical knowledge corpus of 12,500 documents demonstrates significant improvements over baseline GPT-3.5: hallucination rate reduced from 31% to 9% (71% reduction), accuracy improved from 68.4% to 91.2%, and average relevance score increased from 0.72 to 0.94. Response latency remained practical at 2.3±0.4 seconds. The framework achieves F1-score of 0.89, demonstrating robust performance across diverse query types. These results validate RAG as an effective approach for grounding LLM outputs in verifiable knowledge, with significant implications for deploying trustworthy AI systems in engineering and technical domains.
Keywords:
Retrieval-Augmented Generation; Large Language Models; Hallucination Mitigation; Feiss; Lang Chain; Vector Database; Semantic Search
Full text article in PDF:
Copyright information:
Copyright © 2026 Author(s) retain the copyright of this article. This article is published under the terms of the Creative Commons Attribution Liscense 4.0