AI maintenance knowledge search.

What maintenance knowledge search does

AI maintenance knowledge search is retrieval with a short written answer on top. A technician types or says “Line 2 shrink wrapper, fault 214 after a film change” and gets the troubleshooting section for that wrapper’s model, the past work orders where fault 214 was cleared, the parts used, and a two-paragraph summary that cites each source. The value is in finding the right page and the right job card in seconds, especially at 2am when the senior fitter is asleep.

It is not a diagnosis engine. The technician tests and decides. The assistant’s job is to make sure they start from the plant’s own history instead of a generic internet answer. Under the hood this is retrieval augmented generation, and most of the work is in the documents, not the model.

How the workflow runs

Inputs

OEM manuals (often scanned, sometimes translated), the CMMS work order history with free-text notes, site SOPs, parts lists and drawings, shift handover logs, and the asset register with model and serial numbers.

Systems

The CMMS (read-only), a document store such as SharePoint, a private search index tagged by asset, a model API, and a phone or tablet interface usable with gloves.

Steps

Extract text and page images from each manual; split by section; tag every chunk with asset, model variant and document version; expand work order shorthand (“repl brg DE”); retrieve for the named asset only; draft an answer with citations; show source pages alongside.

Human review

The maintenance planner owns the document set and approves new sources. An engineer reviews a weekly sample of answers and every answer a technician marks wrong.

Outputs

Cited answers, links to controlled procedures, and a log of questions with no good answer, which becomes the documentation backlog.

Measures

Retrieval hit rate on a fixed test set of past faults, answers marked useful, time from breakdown call to first action, and repeat faults on the same asset.

Risks

Wrong model variant, superseded procedures, invented values such as torque settings, and answers mixing two machines.

The safety rule: link the procedure, never rewrite it

WorkSafe’s guidance on machine lockouts sets out a lockout process for servicing, which it defines to include maintenance, repair, cleaning and unjamming: identify every energy source, isolate, lock, tag and test before work. Your site’s isolation procedures are controlled documents for that reason. The assistant must never summarise, reorder or “simplify” them. When a question touches isolation, the answer opens with a link to the controlled procedure for that asset and stops there.

This follows from the primary duty in the Health and Safety at Work Act 2015: a business must ensure worker health and safety so far as is reasonably practicable. A search tool that paraphrases a lockout sequence and drops a step is a new hazard. The Act also requires you to engage workers when proposing changes that may affect their health or safety, so bring the technicians in early.

Illustrative example: a Palmerston North packaging plant

Illustrative example, synthetic numbers: a food packaging plant with 40 critical assets, about 3,000 pages of manuals and 11,000 work orders over six years, maintained by 14 technicians across three shifts. Assumptions: 60 breakdowns a month; finding the right information takes 25 minutes on average today and 8 minutes with search; technicians also ask 400 routine questions a month.

The fee line assumes 12,000 input tokens per query (retrieved pages plus instructions) and 600 output tokens on Claude Sonnet 5.5 at US$2 and US$10 per million tokens (Anthropic), and 2.1 million tokens embedded once with OpenAI’s text-embedding-3-small at US$0.02 per million (OpenAI), both as listed on 19 October 2026. The real cost is two to three weeks of a planner’s time cleaning, tagging and checking documents. Downtime minutes saved on a bottleneck line usually justify that on their own.

What stays with a person

Technicians decide what to test and what to replace. Engineers decide changes to preventive maintenance schedules, setpoints and spares holdings. Isolation, permits to work and sign-off of safety-critical repairs follow the site’s existing procedures with competent people. Nobody lets the assistant raise or close work orders unaided.

Where maintenance search goes wrong

• Variant mix-ups. A Mk II and Mk III of the same machine share 80% of a manual and differ on the 20% that matters. Filter retrieval by asset tag before ranking.
• Superseded documents. An old SOP still sits in a shared folder. Index only controlled versions and show the revision on every citation.
• Invented numbers. Torque values, clearances and pressures must come from a visible table, never from model memory. Show the page image.
• Tacit fixes. The best fix is often in a fitter’s head, not a work order. The “no good answer” log tells you what to capture.
• No measurement. Build a test set of 50 past faults with known correct sources and rerun it after every change; see RAG retrieval evaluation.

Scanned manuals raise the choice covered in OCR versus vision models, and access control is covered in building private AI search. The same document discipline helps quality inspection report triage and supplier document extraction. See the manufacturing overview for how they fit together.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.