AI assistant
What it does
The assistant is a panel inside Phasor. It has access to the same model and the same results that you have. For example, ask why a bus is at 0.93 pu. The assistant examines the network, reads the related study output, and answers with the numbers that it read. The answer also lists the elements that the assistant examined. You can compare its reasoning with the model. You do not have to trust it without a check.
The assistant can:
- Read the model. Elements, connectivity, ratings, settings and scenarios.
- Read the results. Each completed study, including the violation list.
- Run studies. It can set up and run a power flow, a short circuit study or another case, and report the result.
- Change the model. Each change goes through the same command gate as your own edits. Because of this, each change is a normal edit that you can undo.
- Explain with citations. When a number comes from a standard, the assistant names the clause. It does not state the figure without a source.
Bring your own key
The assistant uses your own API key, with the provider that you select — Anthropic or OpenAI. Phasor operates no inference service. The application has no AI subscription.
To set up the assistant:
- Select Settings → Assistant.
- Select a provider.
- Paste your API key. Phasor stores the key in the credential store of your operating system, per provider. The key is not in the project file. A project that you share carries no credentials.
- Select a model. A larger model reasons better about a network. A smaller model costs less for routine questions.
Requests go directly from your computer to the provider that you configured. A request includes model data only when the assistant needs that data for the answer. When the panel is closed, Phasor sends nothing.
A consumer subscription is not an API key
The assistant needs an API key from the developer console of the provider. A ChatGPT or Claude consumer subscription does not give access to third-party applications. Phasor does not try to use one.
How to get good answers
- Ask about something specific. "Why is the voltage at Ganta 11 kV low at peak?" works better than "check my network".
- Read the list of what it examined. Each answer lists the elements and the results that the assistant used. Check that list first.
- Let it run a study when the result does not exist. A new run is better than an answer from an old run.
- Review each proposed change before you accept it. A model edit from the assistant appears in the undo history, like each other edit.
Inputs and outputs
| Input | Notes |
|---|---|
| Your question | Plain language. Use the element names that you gave. This helps the assistant find them. |
| Model and results | The assistant reads them when it needs them. Phasor does not upload the full project. |
| Attachments | Files that you attach to a message — for example, a manufacturer datasheet. |
| Provider and key | Per provider, from the credential store of your operating system. |
| Output | Notes |
|---|---|
| Answer | With the numbers that it read, and the elements that it read them from. |
| Citations | The standard and the clause, when a figure comes from one. |
| Proposed edits | As commands. You can review them and undo them. |
| Study runs | Real runs on your model, not estimates. |
Notes
It is an assistant, not an approver. You can check each number that it reports against the model and against the governing standard. Do this check. The assistant makes the path to an answer shorter. It does not sign the answer off.
It knows only what is in the model. If nobody entered the impedance of a cable, the assistant cannot calculate with it. It tells you that the value is missing. It does not assume a typical value.
Your own scripts can use the same tools as the assistant. See MCP automation.