BacLens uses an AI model for four jobs: understanding point names (which of 300 points is the AHU-3 supply air temperature?), drawing system graphics, writing alarm rules, and writing Python control scripts for the simulator or the real building. You choose the model in Settings. This guide explains why the choice matters, where a free local model is enough, and when a stronger one pays for itself.
Why the model changes the result
None of these jobs is about writing text. They are about reasoning on messy site data:
- Point names are cryptic.
AH3_SAT,CTA-03 TEMP SOUF,AV:117 SPLY: the model has to recognise the equipment, the role and the units, often in two languages, and keep it consistent across hundreds of points. - Equipment is connected. A VAV belongs to an AHU, a pump to a heating loop, a valve to a coil. Getting the graphic right means getting those relationships right.
- Control code must be correct, not just plausible. A script that releases the wrong priority or compares °F with °C runs without an error and does the wrong thing.
A stronger model keeps more context in mind, follows the rules it was given more reliably, and makes fewer confident mistakes. On a 20-point demo every model looks good; on a 3,000-point hospital the difference shows.
Start free and private: Ollama
Ollama runs open models on your own PC. In BacLens: Settings → AI → Provider: Ollama, then pick a model.
Good for: - trying BacLens AI features at no cost; - sites where nothing may leave the building network (nothing is sent to the internet); - small mapping jobs and short scripts.
Limits to know: - it needs a reasonably strong PC (more memory = bigger, better models); - small models make more mapping mistakes and lose track on long point lists; - describing a photo of a graphic needs a model that reads images.
BacLens adapts to this: with a small local model it gives the AI simpler tools, one small step at a time.
When to use a stronger model
Switch to a stronger model (BacLens supports Claude through your own Anthropic API key) when:
- the site is large or the names are inconsistent;
- you want graphics reproduced from a photo or a screenshot of the old BMS;
- you ask for multi-step control scripts (load shedding, schedules, sequences);
- a mistake costs more than the API call (it usually does).
You pay the AI provider per use, directly; BacLens does not add anything on top.
Whatever the model: check before it acts
AI output is a proposal, not a decision. BacLens is built around that:
- the mapping and graphics show what the AI proposed, and you correct them;
- scripts written by the AI are checked before they are shown, then run in test mode (they only log what they would write) and in a dry run that lists every write;
- nothing writes to a real controller until you switch it to live.
Summary
| Job | Free local model (Ollama) | Stronger model (Claude) |
|---|---|---|
| Small site point mapping | Usually fine | Fine |
| Large site, cryptic names | More corrections needed | Fewer mistakes |
| Graphic from a photo | Needs an image model, simple results | Best results |
| Short scripts | Fine, check them | Fine |
| Multi-step control logic | Risky, test carefully | Recommended, still test |
Start with Ollama to learn the tools for free, keep it where privacy rules, and use a stronger model where accuracy matters. In every case, let the test mode and the dry run show you what will happen before it happens.