Controlled Industrial Intelligence · Engineering example

Approved Document Search Assistant

Help authorized personnel locate the approved procedure, manual, or maintenance record for the correct asset and revision. Show the source and its status with every answer, and stop when the evidence is missing or conflicts.

The operating challenge

The wrong revision can look like a good answer.

Industrial teams work with manuals, approved procedures, change notices, parts lists, and maintenance histories scattered across document stores. A search result may describe the right type of machine but the wrong serial range, retrofit, site, or effective date. A fluent summary is not trustworthy unless the underlying document is current, authorized for the user, and applicable to the specific question.

Documents have status

A superseded draft or withdrawn instruction must not appear as an active procedure merely because its wording closely matches the query.

Assets have variants

Model, serial range, installed modifications, and site configuration can change which document applies. A broad text match cannot resolve that alone.

Answers need evidence

A generated explanation should point to a stable source, revision, section, and effective status. If no approved passage supports it, the system should say so.

Why modernize: Reduce time spent searching and comparing records while making document authority more visible. This is a governed information workflow, not an autonomous operating or maintenance decision.

Illustrative operating example

A technician asks about a compressor after a controls retrofit.

This compressor-document situation is illustrative. It is not a claim about a completed customer installation, repair outcome, or guaranteed answer accuracy.

Imagine a facility with three similar compressors. Compressor C-2 received a controller retrofit last year. A technician needs the approved inspection guidance for a recurring temperature alarm and sees a search result from an older manufacturer manual. The old manual is genuine but describes the pre-retrofit wiring and alarm mapping. A general-purpose chatbot might merge the old instruction with a newer change notice and produce a convincing composite procedure. A controlled assistant should instead identify C-2, compare the effective revision and retrofit record, cite only applicable approved material, and flag any unresolved conflict for engineering review.

Resolve identity first.Ask for the asset tag and confirm site, model, serial range, controller version, and installed retrofit. Do not guess from an alarm phrase alone.
Filter before retrieval.Apply user authorization, document approval status, applicability, effective date, and version rules before ranking text passages.
Show the evidence.Present relevant sections with title, revision, approval state, and direct document location. Clearly label maintenance history as experience, not an approved procedure.
Expose the conflict.If an approved change notice overrides a manual section, show both and the authoritative precedence rule. If precedence is unclear, stop short of procedural advice.
Keep the decision with people.The technician and responsible engineer consult site procedures and decide what work is permitted. The assistant cannot authorize service, isolation, setpoint changes, or return to service.

What the reviewer would see

Compressor C-2: source-backed answer

The assistant makes applicability and uncertainty visible before a person acts.

Question context: Asset C-2, site, serial range, controller retrofit, and authorized user role.
Approved source: Document title, effective revision, approval status, section, and stable link.
Excluded source: Superseded manual section marked historical and not used as current instruction.
Decision boundary: Conflicting or absent authority routes to document control or engineering.

Other high-priority situations

Search can help without blurring authority.

Shift handoff asks what changed

Summarize the approved change notice and related work record for a named line and effective date. Preserve the difference between an instruction, a maintenance observation, and an unresolved question.

Useful report: cited changes, affected assets, acknowledged readers, and outstanding questions.

Repeated fault history is hard to find

Search technician work notes for similar verified faults and repairs while labeling them as historical cases. Do not promote a past workaround into an approved procedure.

Useful report: linked work orders, verified failure codes, document gaps, and engineering review requests.

What a modernization would connect

Govern the corpus before adding a conversational layer.

The hard problem is not producing prose. It is establishing which documents are approved, who may see them, what assets they cover, and when they became effective. Retrieval and AI must operate inside those constraints rather than treating search relevance as authority.

Required operating evidence

Canonical document ID, owner, approval state, revision, effective date, and withdrawal date
Asset model, serial range, site, installed modification, and applicability rules
Change notices and explicit precedence over older documents
User role and document-level access restrictions
Stable citations to document section or page
Query, retrieved sources, answer version, and reviewer feedback

Document history can be searchable for authorized users while remaining visibly separate from current instructions.

Connected technical layers

Document-control repository as the authoritative source
Asset registry and approved change-control records
Index that carries version, applicability, and access metadata
Retrieval filter and citation verifier before answer presentation
Human review queue for conflicts, stale records, and corpus gaps

The assistant has no connection to the equipment-control path and cannot write an approval back to document control.

Asset and questionAccess and version filterApplicable passagesCited answer or abstentionHuman decision

If asset identity is uncertain, approval status is missing, or applicable sources disagree, return the evidence and an explicit uncertainty rather than synthesizing a new instruction.

Where automation and AI fit

Use AI to navigate evidence, not to create authority.

Deterministic controls

Enforce identity, permissions, approval state, applicability, and revision precedence before retrieval. Reject a draft or withdrawn record even if its text is a strong match.

Retrieval and generation

Find relevant passages and compose a concise, cited comparison. Verify that each claim is supported by the cited passage and allow the assistant to abstain.

Document-quality intelligence

Repeated unanswered queries or conflicting revisions can identify documentation debt for the document owner. The model may recommend review, but only a designated approver changes document status.

Human operating authority

Technicians, supervisors, document control, and engineering retain responsibility for procedures, training, safety, and work authorization.

How we examine a proposed solution

AI advisors challenge source quality and answer boundaries.

These Adaptiv Stratum AI advisory roles can support bounded design review. They are not the document approvers or a substitute for a qualified technician.

Argus

Tests document IDs, version lineage, asset tags, and citation links.

Hephaestus

Checks whether asset variants and retrofits affect the technical answer.

Pythia

Evaluates retrieval and abstention against representative questions.

Aegis

Probes misleading citations, stale revisions, and unsupported synthesis.

Clio

Defines a measured pilot for finding time, citation accuracy, and escalation quality.

Themis

Reviews access boundaries, document authority, and accountability.

What becomes visible

Measure whether users reach the right approved source.

A usage count does not tell whether the assistant is helpful. Reports should reveal applicability decisions, citation quality, abstentions, document gaps, and feedback from the people who actually use the information.

Per-answer record

Question, asset context, user role, and retrieval time
Document ID, revision, approval state, section, and citation link
Excluded or conflicting sources and why they were not used
Answer, confidence limits, and abstention reason
Human feedback, correction, and escalation outcome

Program view

Correct approved-source retrieval on a reviewed question set
Unsupported claims and broken citation rate
Stale, withdrawn, or unauthorized source exposure
Abstentions caused by missing or conflicting authority
Time to find a document and user-reported usefulness
Document-owner backlog and revision-resolution time

A rising abstention rate may reveal a document-control problem, not a model problem. Resolve source authority before relaxing answer rules.

Operating safeguards

Make the boundary visible at every answer.

Current approval and applicability are checked before ranking
Role-based access filters apply before retrieval and response
Stable citation identifies exact revision and section
Historical maintenance notes are labeled as experience, not procedure
Unsupported or conflicting answers abstain and route for review
Revocation and document updates invalidate stale index entries
Adversarial tests cover prompt injection inside retrieved documents
No AI-issued operating instruction, equipment command, or service approval

A practical first phase

Start with one equipment family and a reviewed question set.

Choose the corpus.Use one asset family with approved manuals, retrofit notices, and an accountable document owner.
Map authority.Record approval, revision, effective period, serial applicability, access, and precedence. Remove ambiguous documents from the answerable set.
Build a test set.Technicians and engineers write real questions, including wrong-revision traps, missing answers, and contradictory sources.
Run read-only trials.Compare cited retrieval with manual lookup. Score source validity, claim support, abstention, access control, and usefulness.
Expand only after review.Fix document gaps and controls, train users on limitations, and continue sampling answers after release.

Begin with a document question

Can your team find the effective instruction for a specific asset?

A small, approved corpus can show whether better retrieval or better document control is the first need.

Discuss document search →

Evidence and further reading

Ground generated answers in governed sources.

These NIST sources inform the method. The compressor example is illustrative, not an Adaptiv Stratum customer case study.

NIST: Generative AI Profile addresses risks such as confabulation and the need to document provenance, context, and evaluation.
NIST: When a Manufacturer Asks How Do We Get Smart? shows why unstructured maintenance work orders contain useful but difficult-to-retrieve history.
NIST: Building Evaluation Probes into Agentic AI describes evaluating whether cited claims are grounded in their source material.