Hallucination Detection Framework for RAG Systems

This paper presented a claim-level framework for detecting and verifying hallucinations in RAG systems. The main idea is to check individual factual claims rather than treating the entire generated response as one unit. The framework combines semantic similarity and Natural Language Inference (NLI) to improve verification and uses different outcome categories to distinguish between supported claims, contradictions, partial support, and cases where sufficient evidence is not available. When the internal knowledge base cannot provide enough information, the system can also use external web sources based on their reliability.   The proposed system is designed using locally hosted and open-source technologies, making it suitable for reproducible research without relying on paid APIs. An ablation-based evaluation plan has also been defined to compare the individual verification components with the complete framework. Confidence calibration is included to check whether the confidence values produced by the system match its actual results.   At the current stage, the system is still under implementation and the planned experiments have not yet been completed. Therefore, this work does not claim measured improvements in accuracy or hallucination detection. The next step is to complete the implementation, conduct the planned evaluation, and analyze both the strengths and limitations of the proposed approach based on experimental results.

[1] P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. Küttler, M. Lewis, W. Yih, T. Rocktäschel, S. Riedel, and D. Kiela, “Retrieval-augmented generation for knowledge-intensive NLP tasks,” in Proc. NeurIPS, 2020. [2] C. Niu et al., “RAGTruth: A hallucination corpus for developing trustworthy retrieval-augmented language models,” in Proc. ACL, 2024. [3] J. Li, X. Cheng, W. X. Zhao, J. Nie, and J. Wen, “HaluEval: A large-scale hallucination evaluation benchmark for large language models,” in Proc. EMNLP, 2023. [4] MultiHal and related 2025 studies on semantic similarity, NLI, and embedding limits for RAG hallucination detection. [5] “Rethinking evaluation for LLM hallucination detection,” preprint, 2026. [6] Cleanlab, “Notes on annotation quality in public hallucination benchmarks,” 2024. [7] A. Williams, N. Nangia, and S. Bowman, “A broad-coverage challenge corpus for sentence understanding through inference,” in Proc. NAACL-HLT, 2018. [8] N. Reimers and I. Gurevych, “Sentence-BERT: Sentence embeddings using Siamese BERT-networks,” in Proc. EMNLP-IJCNLP, 2019. [9] Á. Kovács and G. Recski, “LettuceDetect: A hallucination detection framework for RAG applications,” 2025. [10]  Ragas project documentation.