Deploying a Fully On-Premises Retrieval-Augmented Assistant over Proprietary Equipment Documentation in a Manufacturing SME

Abstract

Small and medium-sized manufacturers hold much of their operational knowledge in proprietary equipment documentation — service manuals, standard operating procedures (SOPs), and maintenance bulletins — that original equipment manufacturers distribute under terms that discourage or prohibit uploading to third-party cloud services. This constraint rules out the hosted assistants that dominate current practice. We report on a five-phase pilot in which a manufacturing company in Puerto Rico deployed a fully on-premises retrieval-augmented generation (RAG) assistant over several thousand pages of such documentation, using only open-weight models and open-source infrastructure on a single GPU workstation. We describe the requirements elicitation, procurement, installation, integration, and testing phases, and report an outcome we believe generalizes beyond this site: every phase the project plan treated as risky — hardware selection, model installation, interface integration, baseline model behavior — succeeded on the first attempt, while the step the plan treated as routine — loading the customer’s own documents into a knowledge base — failed outright and consumed more engineering effort than the five preceding phases combined. We trace the failure to three compounding defects in the default configuration and describe the remediation: replacing the default embedding model and adding a vision model to the conversion path, each after a four-candidate comparison, revising the document conversion and chunking configuration, and rewriting the system prompt so that retrieval is mandatory rather than discretionary. We close with eleven lessons for practitioners planning comparable deployments, and report what the remedies did not fix: making retrieval mandatory in the system prompt reduced discretionary retrieval without eliminating it. The system is complete and in daily use, one assistant per machine. This is an experience report, not a controlled evaluation; all outcome assessments are qualitative and we state the resulting limits on inference explicitly.