RAG for Technical Documentation
Helping factory-floor and maintenance personnel find grounded answers about electronics defects in the IPC-A-610F standard.
My work: Built an experimental framework to compare document parsing, retrieval, and answer quality.
Finding: A simpler retrieval setup without a reranker performed best in the evaluated configurations.
Technical details: RAG for Technical Documentation
Design and experiments
The framework covers layout-aware parsing, structural and asset validation, synthetic dataset generation, dense, sparse and hybrid retrieval, reranking, generation models, prompt strategies, and evaluation with Ragas and Langfuse.
What the comparisons showed
Docling produced the most robust representation in the experiments. The best retrieval configuration used text-embedding-3-small, a three-chunk window, and no reranker. Domain-role framing removed refusal-style answers in the evaluated query set, while explicit instructions improved figure-reference retention.
Evaluation scope
The final configuration was checked against a separate 30% synthetic holdout set: questions reserved for the final evaluation. These findings describe the evaluated document and query set; they do not establish reliability for other documents or real factory use.
System pipelines
Pipeline 1: compare document parsers, build a structured graph, generate evaluation datasets, and select the most robust parser.
Open full parser-selection diagram
Pipeline 2: optimize retrieval, generation, and prompts in sequence, then validate the final RAG configuration on held-out questions.
Open full RAG-optimization diagram