On-Site Diagnostic Reasoning Assistant

An on-site diagnostic reasoning assistant retrieves manuals, wiring diagrams, prior tickets, and parts notes for the asset in front of the technician, then ranks plausible root causes and a repair sequence. The goal is shorter time to a confident first action on the floor or at the customer site, not an unsupervised close.

The assistant answers questions such as: which failure modes match these symptoms for this model, which checks come next, and which torque, clearance, or safety notes apply. The technician still owns the diagnosis. They run the checks, verify the outcome, and record what worked.

Speed comes from cutting search and guesswork. A lead with dozens of SKUs and thin tribal knowledge gains the most when the corpus is model-complete and ticket history is clean enough to surface repeat failures.

This pattern fits discrete manufacturing and industrial equipment where service is sold against a defined installed base. Typical signals: multi-hour average repair time, high first-visit incomplete rates driven by wrong parts or wrong sequence, and technicians who spend more time hunting PDFs than turning wrenches.

It also fits mixed fleets where OEM manuals, dealer bulletins, and internal SOPs disagree. Ranking across those sources reduces “which PDF is current” debates during a live call.

Keep the feature empty, or return a clear no-corpus state, when the asset model has no usable manual set, diagram pack, or ticket cluster. Inventing steps from a neighboring model is worse than a blank card. Same rule for prototypes, custom one-offs, and assets whose documentation lives only in a supplier portal your retrieval pipeline cannot reach.

Gate activation on asset identity: serial, model, revision, and software/firmware where it changes procedures. Without a reliable identity match, do not rank causes. Surface “manual corpus missing for this model” and fall back to standard work or a senior callback.

Retrieval-augmented generation (RAG) over approved manuals, exploded views, and closed tickets is the core. Ranked root causes should cite the passage, figure, or ticket that supports each hypothesis. Repair sequences should list ordered checks with tools, measurements, and stop conditions, not a prose essay.

Prefer short ranked lists over long narratives. Three to five causes with evidence beats a single confident paragraph. Call out safety interlocks, lockout/tagout, and torque or pressure limits in the first screen of the sequence.

The technician confirms the fix before the system treats the job as solved. Confirmation can be a measured value in range, a successful self-test, or an explicit “cause verified” with parts used. Until then, keep alternate causes visible so a wrong first pick does not bury the next best option.

Offline and low-connectivity sites need a downloaded model pack or cached manual slice for common assets. Sync ticket embeddings when back online so the next visit benefits from what this visit found. Do not require a live chat session to show the top causes for a known model.

Field service platforms such as ServiceMax already hold job context, asset history, and parts. Pairing ranking with that context avoids a second app for identity and work order status. PTC ecosystems often already store CAD, service manuals, and digital twin or product data that enrich diagram-linked retrieval. Microsoft stacks (Azure AI Search, Copilot Studio patterns, or Fabric-connected knowledge) suit plants that want corpus governance, identity, and logging inside an existing Entra and Purview boundary.

Choose the hub by where the work order and asset master live. Put retrieval and ranking next to that system of record. Avoid a free-floating chatbot that cannot see which asset is on the work order or which revision the unit carries.

Whatever the stack, keep an audit trail: query, retrieved chunks, ranked causes, technician selection, and final confirmed cause. That trail feeds warranty and continuous improvement later. It also lets engineering spot manuals that never get cited because they are unreadable or outdated.

Define the corpus ownership model before pilots. Service engineering owns procedure accuracy; product documentation owns OEM manuals; field ops owns ticket quality and close codes. Without owners, RAG quietly ranks stale bulletins.

Start with a narrow model family where manuals are digitized, diagrams are indexed, and closed tickets use consistent failure codes. Measure first-visit fix rate, time-to-first-correct-action, and “no corpus” rate. Expand only after identity matching and citation quality hold under live load.

Train technicians to treat ranks as a checklist order, not a verdict. Escalation to a senior should remain one tap when symptoms fall outside the ranked set or when two top causes conflict. Capture overrides: when a tech rejects the top cause and succeeds with another, feed that outcome into ranking and into the manual backlog.

Pair this assistant with upstream and downstream work. Dispatch that sends the right skill and parts reduces empty visits before diagnosis starts. Health signals and failure prediction can pre-load likely causes when the asset already showed precursors. Warranty classification benefits when the confirmed cause and cited procedure are structured on the close.

Wrong model match is the highest-risk failure. A near-SKU manual that looks right can authorize the wrong torque or omit a recall. Hard-fail when confidence on identity is low.

Hallucinated steps without citations are the next risk. Require every actionable step to point at a manual section, diagram callout, or prior ticket. If retrieval returns nothing useful, show empty state, not a generated guess.

Ticket pollution also hurts ranking. Uncoded closes, copy-paste notes, and “fixed / unknown” outcomes teach the system noise. Clean close codes and a short verified-cause field matter as much as the LLM.

Finally, do not optimize for chat fluency. Optimize for a technician who has gloves on and five minutes before the next site. Dense, cited, ordered guidance with a clear empty state beats a conversational agent that cannot say “we have no manual for this unit.”