Key facts
| API surface | OpenAI-compatible /v1/chat/completions; keep existing SDK tooling |
| Data grounding | Embeddings and RAG are live for manuals, SOPs and maintenance history |
| Agents | Function calling and agent orchestration are live for MES and ERP lookups |
| Structured output | JSON mode returns structured quality and work-order records |
| Deployment | Cloud, VPC, on-prem or air-gapped inside the plant network |
| Models | 30+ models including long-context options for large manuals |
| Pricing model | Flat monthly self-serve plans; no per-token billing on self-serve |
| Endpoint roadmap | Images, audio, moderation, batch and fine-tuning are coming soon |
TL;DR
- Knowledge search is the fastest win: manuals, SOPs and maintenance history with citations.
- Agents read MES, ERP and ticketing data through APIs before any write access is granted.
- Vision inspection is on the roadmap; text and embedding pipelines are live now.
- Air-gapped deployment keeps plant and process data inside the network.
- Measure time-to-answer and first-time-fix rates, not just model benchmarks.
How it works, step by step
- Pick one knowledge domain, such as line-specific maintenance or quality procedures.
- Collect and classify documents: manuals, SOPs, work orders, incident reports and supplier specs.
- Choose a deployment target; use on-prem or air-gapped when process data cannot leave the network.
- Chunk and embed documents, then test retrieval with real troubleshooting questions.
- Expose read-only MES, ERP or historian APIs as tools and confirm permission enforcement.
- Return structured records in JSON mode and route them into existing work-order queues.
- Expand by adding document classes, then evaluate vision inspection once endpoints are available.
Try it yourself
Open the best model for RAG selector →
Knowledge search is the fastest win
Maintenance and process knowledge is usually scattered across PDFs, spreadsheets and the memory of senior staff. A retrieval assistant changes that quickly:
- Maintenance troubleshooting: search fault codes, past work orders and service bulletins with citations.
- SOP retrieval: answer line and safety procedure questions from the current approved revision.
- Quality reporting: summarize defect reports and surface recurring causes for engineering review.
- Supplier documentation: check certificates and specifications against required fields.
Answers cite their source, so technicians can verify before acting. Keep the source visible in the interface.
Agents on the plant floor: read first
Give an agent read-only tools before any write capability: work-order lookup, machine status, parts inventory and shift schedules. Function calling is live, and your service layer decides what the model can see, based on the requesting user's permissions.
Write tools such as creating a work order should require an approval step and produce a logged, reversible action. That keeps automation auditable when production pressure is high.
Vision and inspection on the roadmap
Dedicated image and vision endpoints are coming soon, so plan the pipeline now: camera capture, frame selection, labelling strategy and a human review station for low-confidence results. Traditional vision systems and upstream OCR can still feed text into the live API today - for example, extracting gauge readings or printed labels, then checking them against specifications with JSON mode.
Running inside the plant network
Process data, supplier terms and production volumes are sensitive. Plugsky supports cloud, VPC, on-prem and air-gapped deployment, and the API shape stays OpenAI-compatible across modes. That means a pilot built on public manuals can be redeployed inside the plant network without rewriting the application.
Honest comparison
| Capability | Plugsky | Traditional enterprise search | Building in-house |
|---|---|---|---|
| Manuals and SOPs | RAG with citations over approved documents | Keyword search, no synthesis | You build ingestion and evals |
| Agent actions | Function calling into MES, ERP and ticketing | Not available | Custom orchestration |
| Vision inspection | Coming soon | Rule-based machine vision only | Separate ML pipeline |
| Deployment | Cloud, VPC, on-prem, air-gapped | On-prem only | Your infrastructure |
| Model choice | 30+ models behind one API | Not applicable | You host each model |
| Pricing | Flat monthly self-serve plans; see live pricing | License and maintenance fees | GPU plus operations cost |
Frequently asked questions
Can we run this inside the factory network?
Yes. On-prem and air-gapped deployments keep prompts, documents and embeddings inside your network, with no outbound connectivity required.
Is image-based inspection available?
Dedicated vision and image endpoints are coming soon. Today you can run OCR or upstream vision tooling, then send the extracted text to the live chat and JSON mode endpoints.
How do we connect MES and ERP?
Expose narrow, permissioned APIs and register them as tools for function calling. Start read-only and log every call with arguments and results.
What stops wrong maintenance advice?
Answers must cite approved manuals or work orders, uncertain cases should escalate to a senior technician, and safety-critical procedures always require human confirmation.
Which model handles long manuals?
Long-context models in the 30+ model catalogue help with whole-document questions. Combine them with chunked retrieval so citations remain precise.
How do we evaluate the pilot?
Track time-to-answer, first-time-fix rate, escalation rate and technician corrections on a labelled set of real questions.
Can we start without a paid plan?
Yes. The free plan includes two free AI models with no card, and the 14-day full-access trial lets you test larger models against your own documents.