Turning maintenance documents into answers people can verify.

01

The information existed, but the answer did not

The fictional service business Meridian Equipment supported multiple generations of industrial machinery. Manuals, change notices, and internal troubleshooting notes were stored in different systems. A technician searching for a fault code frequently found the right phrase in the wrong equipment revision.

The scenario goal was to reduce document-finding effort while preserving the technician’s judgment. It was not to let an assistant authorize maintenance work. Equipment isolation and safety procedures continued to follow the established operational process.

02

Structure the corpus around equipment identity

We first normalized equipment families, document revisions, and publication status. Superseded documents stayed discoverable for investigation but were excluded from ordinary answers unless explicitly requested. Every indexed passage retained a link to its source and the access rules attached to the document.

The search layer combined exact terms with semantic retrieval. Fault codes and part identifiers benefited from literal matching; descriptions such as “intermittent vibration after warm-up” benefited from semantic matching. The system merged these candidates, reranked them, and presented a small set of inspectable sources.

03

An assistant that can say there is not enough evidence

A response had to identify the relevant equipment revision and cite the passage behind its recommendation. If the sources disagreed or the equipment identity was missing, the assistant asked for clarification instead of composing a complete-looking procedure.

We created a reviewed question set covering ordinary lookups, outdated manuals, ambiguous component names, and permission boundaries. The same set was rerun after ingestion changes, prompt changes, and model upgrades.

• Test search relevance separately from answer quality.
• Preserve document-level authorization during indexing and retrieval.
• Return source passages with equipment and revision context.
• Include missing and contradictory evidence in the evaluation set.

04

Illustrative outcomes, with a stable question set

These invented measurements compare 240 reviewed questions in the scenario. A supported answer means the cited source answered the question for the correct equipment revision; it does not mean a maintenance action was independently safe.

05

A reusable pattern for knowledge-heavy operations

The architecture is useful wherever information has versions, owners, and consequences: support, field service, compliance operations, or internal engineering. The difficult work is usually source hygiene and evaluation, not the final chat interface.

The scenario handover included ingestion ownership, source-correction procedures, and a weekly review of unanswered questions. That keeps a search service useful after the initial launch rather than turning it into a polished interface over an aging corpus.