Reducing Hallucinations in Large Language Models Through Integrated Self-Verification and Retrieval-Augmented Generation

Large Language Models (LLMs) are increasingly deployed for advanced engineering tasks such as Computer-Aided Design (CAD) documentation, standards compliance verification, and knowledge retrieval. However, their tendency to generate hallucinations—convincing but unfounded outputs—undermines trustworthiness in high-stakes engineering applications where precision and regulatory compliance are paramount. This paper addresses the critical need for reliable LLM outputs by introducing CoVe-RAG+, a unified framework that combines Chain-of-Verification (CoVe) with Retrieval-Augmented Generation (RAG). The goal is to mitigate hallucinations by grounding LLM responses in authoritative external sources, including engineering standards, CAD data, and simulation reports, while employing iterative self-verification to validate key claims. The work is positioned within the broader context of enhancing LLM reliability for engineering design processes, where factual accuracy is non-negotiable.

CoVe-RAG+ operates through a two-pronged approach: retrieval augmentation and iterative self-verification. First, given a user query, the system retrieves relevant documents from a curated knowledge base comprising engineering standards (e.g., ISO, ASME), CAD model metadata, and simulation reports. These retrieved passages serve as external evidence. Second, the LLM generates an initial response, which is then subjected to a Chain-of-Verification process. This involves decomposing the response into atomic claims, formulating verification questions for each claim, and re-querying the retrieval system to confirm or refute each claim. The verification results are aggregated to produce a final answer, along with a verification report that highlights supported claims, unsupported claims, and source traceability. The process can be formalized as follows: let $q$ be the query, $D$ the retrieved documents, and $r_0$ the initial response. For each claim $c_i$ in $r_0$, a verification question $v_i$ is generated, and a retrieval function $\text{Ret}(v_i)$ returns evidence $e_i$. The claim is accepted if $e_i$ entails $c_i$; otherwise, it is revised or rejected. The final response $r^{\ast }$ is constructed from verified claims. This iterative loop continues until all claims are either verified or flagged. The framework is designed to be model-agnostic and scalable, integrating seamlessly with existing LLM pipelines.