# Orenda AI Agent

The Agent runs machine investigations and reports on your OrendaBox. It combines approved machine documents, production and stop history, maintenance records and available PLC telemetry. Each report identifies its evidence, missing data and suggested checks.

Open **AI workspace → Work** to start or continue an investigation. The saved conversation keeps its question, machines, period, evidence, next actions and report versions together. **Knowledge** opens the existing machine documents and reviewed resolutions; **Automations** manages the existing Agent routines. The general chat launcher remains available elsewhere for explanations, but operational investigations belong to this saved workspace.

**Reports** has **AI reports** and **Standard reports**, with **Analytics PRO** available to Pro users. Standard reports retains the existing production views. **Analyze with AI** carries the selected machines and period into a saved AI report. The maintenance board requires MES 0.1.54 and Edge Manager 0.2.78 on Platform 0.2.84 or newer; update MES and the Agent together. Consult the release record for measured latency and remaining deployment qualification.

## Create and refine a report

1. Open **Reports → AI reports** and describe the question, audience and desired result. Select the machines and period, report language and level of detail. Add a supplied file when it provides necessary context.
2. Start the investigation. The conversation shows requests and findings; **Activity** shows tool operations and the public question checklist. You can browse other pages or reload. The header links back to work in progress or an unread result.
3. Review the report and its **Data** sources. Ask for a comparison, a deeper investigation or a presentation change in the same conversation. Existing versions remain available. If an input is missing, supply that specific input; a paused work budget instead offers **Continue** with the collected evidence retained.
4. Use **Versions** to inspect or restore an earlier result. Export the selected result as PDF or Excel. Large workbooks provide a partition manifest identifying included rows and any source omissions. Source permissions and retention still apply when exporting.

The report library supports search, templates, team sharing, archive and deletion. A template copies the report brief into a new report; it does not pass old measurements off as newly collected data. Sharing requires the recipient's current organization and source access. Unsent conversation text is saved, but it does not start paid inference until you send it.

## Continue a maintenance case

1. In **Predictive Maintenance**, select a machine and describe the change you want to investigate. Coverage shows which signals have been assessed; an unassessed signal is not a machine fault. The case opens in the same saved workspace used for its later questions and reports.
2. Record the technician's actual findings, unsuccessful checks and unavailable measurements. Correct an earlier observation through its correction action so both the original and corrected record remain attributable. The next investigation uses this case history alongside authorized source data.
3. Review a proposed action before creating work. Select the equipment, assignee, checklist and verification criteria in the existing MES work-order interface, or link a relevant order already open. A proposal does not operate the machine or create an order by itself.
4. Start and complete the order in MES. Required checklist answers must be supplied; completion is delivered to the case once. Completing an order leaves verification due rather than declaring the repair successful.
5. Investigate the result using comparable periods outside the intervention. The verification tool checks independent runs, product, signal mapping, replacement/calibration history, speed, load, exposure and sample coverage. Missing controls produce an inconclusive result and a specific next check. A measured change still needs a human review of the success criteria before resolving the case.
6. Export the case handover for the next shift, or ask for an executive PDF and supporting Excel workbook in the same conversation. Review a resolution before publishing it for reuse; later contradictory evidence or scope corrections require renewed review.

Draft text is saved on the server. Navigation and reload preserve it; conflicting edits require review instead of silently replacing another editor's work. A failed job retains the previous report and collected evidence. **Continue** or **Retry** resumes the saved request when permitted by current source access and spending allowances. Activity shows public actions and question status, not private model reasoning.

## Run a routine

Create a routine from the existing workspace. Calendar routines refresh its reporting period; condition routines evaluate an approved signal and its configured trigger/recovery thresholds; work-order completion routines follow the linked completed work. Use snooze, suppression, expiry or pause when the investigation should not run. Thresholds are operational settings supplied by an authorized user, not automatically inferred OEM limits.

Condition checks run locally and do not call the model for every sample. A qualifying episode starts a saved investigation subject to current permissions, available worker capacity and AI allowance. The routine retains its delivery state to avoid duplicate investigations for the same episode or completion event.

## Set up

1. Update the Box to Platform 0.2.83 or later with Edge Manager 0.2.77 or later, and install Orenda MES 0.1.52 or later.
2. Install Orenda AI Agent through the Edge app catalog. Review its requested permissions and select the machines, devices, tags and data sources it may access. Empty machine selections do not grant access to every machine.
3. In Edge Manager, enable the MES runtime integration for the selected machines. The Agent and MES must both have permission to use them.
4. Open MES through Edge Console. Add machine details and relevant manuals, maintenance instructions and technical documents. An administrator must approve uploaded documents before investigations can use them.
5. Check AI readiness. The Agent checks the inference profile, its data connection and its offline document tools. Resolve the displayed cause if a check is blocked.

## Hosted and self-hosted inference

Hosted inference uses Orenda's managed service, Box identity and centrally configured limits. You do not need to provide a provider key. The model is fixed by the platform; users cannot choose a model.

A Box administrator can instead configure a self-hosted OpenAI-compatible chat-completions endpoint in Edge's AI settings. Credentials are stored by the host and are never supplied to the Agent or document workers. The endpoint must support the configured model. The Agent itself has no Internet access. Only the host broker sends approved data classes to the configured inference service.

## Run an investigation

Select the machine, time window and optional production context in MES, then describe the question. Useful questions include changes in scrap between product runs, conditions associated with stops, maintenance checks supported by a manual, or differences between comparable production periods.

The job page shows progress, source coverage and any missing permissions or data. Jobs continue outside the browser. Cancellation stops further work; an inference request already dispatched may still consume usage. A failed or interrupted job retains its evidence and progress. Retry follows the displayed recovery action and does not blindly repeat an ambiguous paid inference request.

“Preparing approved access” means Edge is creating the job's scoped access grant.
The analyzer starts after that approval is stored. If approval is interrupted by
a restart, retry the existing job to request fresh access.

Reports include observations, source references, limitations and recommended checks. Open the evidence to inspect the data behind a claim. PDF exports and other generated artifacts are retained under the Box's approved storage allowance and retention policy.

Maintenance investigations automatically screen approved mapped historian gauges alongside archived runs and stop records. The priority list shows measured changes between comparable production runs, with sensor charts, recorded downtime and scrap associations. Product changes alone do not count as deterioration: comparisons stay within the same product and historical sensor mapping, and paused samples are excluded. Missing comparable runs or conflicting sensor sources show “Needs data”. The attention score orders inspection work; it is not a health score or failure probability. Review the cited manual and confirm a technician inspection before creating a maintenance work order in MES.

## Documents and data quality

Approved PDF and DOCX documents are processed locally. Scanned PDFs use OCR and are marked for review because recognition can be wrong. Unsupported, infected, oversized or damaged files produce an actionable failure rather than invented document content.

Every new attachment is scanned before extraction. On a small ARM Box, the first
scan can take over a minute while signatures load. Scanning has a three-minute
limit within the five-minute document processing limit; a timed-out scan blocks
extraction and explains how to retry.

Later investigations reuse your already processed manual when MES still approves
your access and its version and file checksum match. Changed or expired sources
are processed again; access revocation immediately prevents reuse.

Processing a document completes on the Box without using hosted AI allowance.
The result shows extracted sections, source anchors and any OCR review needed.
Start a separate investigation to ask questions about that indexed document.

If the Box is busy with other approved work, wait for capacity and retry the
saved task. Its evidence, draft and conversation remain available. The Agent
reconciles an interrupted inference before attempting another request. A time
or context budget pause offers **Continue anyway** on the same task; hosted
allowances still apply. If an earlier charge is unresolved, the continuation
explains that its reservation remains counted. A paid request is never silently
replayed or removed from usage.

PromQL queries are restricted to approved metrics, device/tag selectors and bounded time windows. A recent Prometheus scrape does not prove that a PLC value was freshly acquired. Historical production correlations use the mappings recorded with a run. Unknown mappings, overlapping ambiguous intervals and continuous quantities with unverified units are excluded from affected claims.

## Recurring investigations

An administrator can create a recurring investigation with a rolling lookback window. It uses an explicitly approved service workload grant, rather than a saved browser session. Every run is bounded and remains subject to current organization access and installed permissions. Removing the approving administrator's access, revoking the grant or changing approved scopes can stop the schedule; the UI explains when another administrator must renew ownership.

Report schedules use the plant timezone and daily or weekly execution time. **Previous complete days** selects complete local dates, including daylight-saving changes. Scheduled runs collect a fresh snapshot and retain earlier report versions. In **Schedules**, pause or edit a recurrence and review its next run. If an app update invalidates its saved access, review and save it again to approve the current app. A failed schedule retains the report and shows the recovery action.

In **AI workspace → Automations**, open saved work to configure its purpose,
calendar or event trigger, minimum source evidence and daily start limit.
Condition triggers can require fresh samples and sustained duration, with
separate recovery thresholds. Work-completion triggers use the recorded MES
completion event. Select result delivery and authorized recipients, then review
the routine's activity, skipped-run reasons and resulting task. Hosted allowance
is visible beside the automation inbox. Preventive maintenance schedules remain
in MES; these routines investigate evidence and propose reviewed follow-up.

## Interpret the result

Open **Predictive Maintenance** to review machine priorities, compare recent measurements with their baseline, inspect source evidence and plan the next maintenance check. Reports distinguish measured findings from possible explanations and link each finding to its sources.

Data availability and pending actions appear beside the relevant operation. Expand report context for source coverage and calculation details. Record the inspection outcome and compare subsequent operating periods to continue the maintenance case.

The **Machine board** starts with the whole authorized fleet, ordered by the
action needed. Select a machine to see the reason, retained measurements,
comparison charts and the next check in the same screen. Missing measurements
are shown separately from a supported fault finding. Use the resize handle or
fullscreen control for a larger view; updates preserve the machine being read.

**Follow-up plan** combines inspections, unanswered investigation questions,
native work orders and verification checks. Open the existing work from its row.
Assign an owner and next check, record the inspection outcome, and compare
subsequent operating periods. A report does not close a maintenance issue.

Open **Investigations**, or the selected machine's investigation, to continue the
same saved conversation. **Findings** and **Orenda AI** stay beside the machine;
navigation and reload preserve the case and its supporting evidence.

Use **Monitoring plan** to schedule daily, weekly, six-hourly or sustained-signal
reviews. Choose current machine signals and a clear threshold for condition
plans. Stable reviews update the board quietly; meaningful changes reuse the
existing issue and surface the next action. Pause or edit the same plan from MES
or AI Workspace. Failed reviews retain evidence for an explicit continuation.

Under **Browse machines → Signals**, choose the measurements and operating
regime in **Signals & regime**. Product, process load, component and warm-up
controls define which recorded runs can be compared. Save a named baseline,
then compare a later period with that unchanged snapshot. Read the included and
excluded-run explanations before interpreting a difference. **History** brings
documents, inspections, component events, work orders, investigations and saved
reports into one timeline; each item opens its source.

In **AI workspace → Work**, start with a question or reopen saved work. Findings,
the next action and the conversation share the same task. Record an unsuccessful
check or corrected observation there, review proposed work through the existing
MES work-order flow, and compare the measured result after completion. A ready
report does not resolve an investigation that still needs evidence. Approved
report versions can be exported to PDF and Excel.

Use **Knowledge** to search authorized machine documents and reviewed knowledge.
Review document edition, model applicability, extraction state and conflicts
before using a source. Indexing runs locally; superseded or withdrawn sources
remain auditable and cannot silently become current instructions.

The MES overview counts the entire currently authorized saved-report library, including older pages. Edge Manager 0.2.68 and MES 0.1.43 enable bounded source authorization batches, so reports sharing sources do not repeat remote access checks. Opening or exporting a report checks its sources again. Changed manuals and revoked sources exclude affected reports; permissions are never cached between requests.

For help, contact [Orenda support](https://orendanet.com/contact).