The assistant

Building a model by hand means placing every component, finding a cost for each one and loading a trace for every renewable. The assistant does that first pass from a description, and then stays around for the edits and the questions. It is on every tier.

What you can ask it

Open Levelised and click the assistant button at the bottom right of the canvas. It works on the model that is open, so start from a blank project or from one of your own.

• Build a model. "A 100 MW mine in the Pilbara, off-grid. What is the cheapest mix of solar, wind, batteries and gas?" It asks for anything it can't default, such as a delivered fuel price, then adds the buses, the load and each supply option, and you watch them appear on the canvas. A build from scratch takes several minutes.
• Change the model. "Add a 4-hour battery" or "make the gas price $14/GJ and re-solve". Every edit marks the results stale, as it would if you made it yourself.
• Understand a result. "Why did it build so little wind?" It answers from the solved model on your canvas and names the components involved, which you can click to select.
• Fix a failed solve. When a solve fails or comes back infeasible, "Ask the assistant" appears on the infeasibility message and in the solve log. It reads the solver log, says the most likely cause, and asks before changing anything.
• Find out how something works. It answers questions about Levelised from this documentation and about the optimisation from the PyPSA user guide, and says which is which.

It asks before it solves. A build ends with a summary of what it put on the canvas and a question; say yes and it runs the year.

What it can't do, and how you do it yourself

The assistant works through a fixed set of actions on your model. It can give a load a profile and build solar and wind traces for a location. It cannot load any other hourly series, such as an hourly electricity price, a time-of-use tariff or an inflow. Hourly values are stored by the app and never pass through the conversation, and the assistant can neither open a file on your computer nor download a data file from the web.

The app takes an hourly series on any input that varies in time, so this is a step you do yourself:

1. Get the data. For wholesale prices in the National Electricity Market, AEMO publishes price and demand files by region and month. Ask the assistant which file you need; it will name the source.
2. Make a CSV with a header row and one row for each snapshot: 8,760 for an hourly year. Name the column Component type.Component name.Attribute, for example Generator.Grid supply.marginal_cost. Main menu → Export Time Series Inputs CSV downloads the series already in your model in this format, which makes a ready template.
3. Main menu → Import Time Series Inputs CSV, then solve again.

Keep the hours on the same basis as the rest of the model. Traces built from a map location are in local standard time for the site, with no daylight saving. AEMO's market time is UTC+10 all year, which matches a site on the east coast as it is.

Once the series is in, the assistant can work with it like any other input: ask it what the prices did to the result.

Where its numbers come from

The assistant is told never to make up a number. Technology costs, efficiencies and lives come from CSIRO GenCost. Emission factors come from the National Greenhouse Accounts Factors. Solar and wind traces are built for the location from the same weather data the map picker uses. Anything site-specific comes from you, or from a public web page that it names.

Each figure is recorded as the source of the input it went into. Open a component and you can see, against each input, where the number came from; for a figure from the web that includes the address and the date. The assistant also says in its reply which figures came from the web.

Health warning: it can still be wrong. GenCost figures are national screening costs, a web page can be out of date, and the assistant can pick the wrong row or misread a request. Click through the components it built and check the cost basis and efficiencies before you rely on the answer. That is the check we run on every build ourselves.

The weekly allowance

The assistant comes with a weekly allowance on every tier. The paid tier has ten times the Demo allowance.

• The bar at the top of the assistant window shows the share of this week's allowance you have used and when it resets.
• It resets every Monday at 00:00 UTC, which the app shows in your local time: 8:00 am in Perth.
• A build that researches figures on the web uses far more of the allowance than a question about a result. The bar moves after each reply.
• The allowance is checked before a reply starts. A reply that has started always finishes, so a build is never cut off part-way.
• You get a notice in the conversation when you pass 80%.

At 100% the message box is replaced by the reset day and a button to contact us: about the paid tier if you are on Demo, or about a larger allowance if you are already paid. Your past conversations stay readable, and the rest of the app carries on as normal.

Your tier's other limits apply to what the assistant builds. Solves it runs count towards your solves per day, and the model has to fit your tier's size and solve-time limits. If a solve looks likely to run past the time limit, the assistant says so before it starts.

What is sent to the AI provider

The assistant runs on an AI model from Anthropic. Each message sends Anthropic your message, the conversation so far, and a description of the open model: its components, their parameters and sources, and summaries of results and time series. Hourly time series are not sent; they stay in Levelised. Your name and email address are not sent.

Anthropic's commercial terms do not allow it to train its models on what we send. The detail is in our privacy policy. If a model must not leave Levelised, don't use the assistant on it; everything else in the app works without it.

Conversations

Conversations are kept with the project they belong to. The list button in the assistant window shows the earlier ones for the open project, and you can delete any of them from there.

Try it

Open a blank project, describe a site you know well, and let it build. Then click the largest generator and check its cost and where it came from against what you would have used.

On the paid tier you can also connect your own AI to Levelised. Back to the documentation.