Domain-Adapted Retrieval-Augmented Generation for Technical Documentation: Enhancing Reliability and Faithfulness in Technical QA

This paper introduces a domain-adapted Retrieval-Augmented Generation (RAG) pipeline specifically designed for technical question answering within industrial documentation. To overcome the limitations of generic retrieval systems in processing version-sensitive and semantically dense content, we develop a structured bilingual corpus of 20,000 manuals and fine-tune a dual-mode embedder (BGE-M3) with triplet loss and hard negatives. We further incorporate a two-stage reranking mechanism and an LLM-based query auditing module (LLM-Critic) to enhance retrieval precision and contextual adequacy. Experimental evaluation demonstrates substantial improvements in Recall@5 (from 0.67 to 0.86) and MRR (from 0.52 to 0.80), confirming the effectiveness of layered retrieval optimization in reducing ambiguity, strengthening grounding, and mitigating hallucinations. We have identified further tasks to address remaining challenges in managing multimodal content, resolving version disambiguation, and ensuring scalability for real-world deployment.