Co-Pilot architecture

The Co-Pilot feels simple — you ask a question, you get an answer. This page explains what happens in between, and why the answers are trustworthy rather than made up.

​

Grounded, not guessing

The key idea: the Co-Pilot doesn’t answer from a language model’s memory. It answers by reading your live data through a fixed set of tools, then putting the results into words. So “how many devices are offline?” returns your actual count right now — not a plausible-sounding number.

​

The tools are the source of truth

Behind the Co-Pilot is a catalog of read-only tools over your data — list devices, batch and project summaries, annotation backlog, OEE and model-performance metrics, and so on. Each tool runs a real, scoped query. The language model’s job is to pick the right tools for your question and phrase the result; the facts come from the tools. This catalog is exposed over MCP (Model Context Protocol), the same open standard that lets external AI clients connect — see Co-Pilot & MCP.

​

The model

The language model that interprets your question and phrases the answer is a compact 7-billion-parameter open-weights model — Qwen2-7B-Instruct by default. The backend is pluggable — it can run on-device on an edge NPU (the default), on a self-hosted Ollama server, or against any OpenAI-compatible endpoint. Because it’s an open model that can run inside your own environment, the Co-Pilot doesn’t have to send your questions or data to a third-party AI provider — a real advantage for closed or air-gapped deployments. And since the facts come from the tools, a compact 7B model is enough: it’s choosing which queries to run and wording the results, not recalling your data from memory.

​

Read-only and org-scoped by design

Two properties make the Co-Pilot safe to hand a question to:

• Read-only — every tool reads; none of them change your data. The Co-Pilot can tell you the annotation backlog, not clear it behind your back.
• Organization-scoped — the tools only ever see your organization’s data, because they run with your identity. There’s no path to another tenant’s information.

​

Streamed, with structure

Answers stream back token by token, so you see them form. Alongside the text, the Co-Pilot returns rich cards (device lists, metric tiles, image galleries) and suggested follow-ups, and it resolves natural time expressions like “last week” or “yesterday” in your timezone. Your conversation history persists, so you can pick a thread back up later.

​

Next steps