TechDoc-RAG: An Evidence-Governed Framework for Software Technical Documentation

Technical support repositories contain large numbers of manuals, technotes, and troubleshooting records, yet their value depends on whether a user can locate and trust the evidence behind an answer.This study develops TechDoc-RAG, an evidence-governed retrieval-augmented generation pipeline for software documentation.Its retrieval layer joins dense semantic search with BM25 lexical search by Reciprocal Rank Fusion (RRF); its response layer separates answerability assessment, corrective retrieval, grounded generation, citation validation, and post-generation support checks.The experiments use NVIDIA TechQA-RAG-Eval and the related TechQA technical-document corpus.On a fixed 100-question retrieval test, Hybrid RRF raised Recall@1 from 0.39 to 0.43, Recall@3 from 0.50 to 0.60, Recall@5 from 0.60 to 0.64, and MRR@5 from 0.4608 to 0.5178 when compared with dense retrieval.The tested cross-encoder reranker added substantial delay without improving the main ranking outcomes.On the separately locked 50-question generation test, the final pipeline increased citation validity from 23.33% to 60% and final refusal accuracy from 0% to 55%, reduced average latency from 1231.26ms to 407.53ms, and eliminated empty outputs.Unsupported outputs decreased from 28 to zero according to the implemented validation criteria.Answer semantic similarity fell from 0.2741 to 0.2356 and the context-support similarity proxy fell from 0.4723 to 0.4111.These findings expose a practical trade-off: stricter evidence controls improve traceability and abstention behavior, but do not improve every automatic semantic measure.