Skip to content

ClimAID assistant (climaid ai)

Under testing

The assistant is new in 0.4.1 and under active testing. It runs the same ClimAID v2 methods as the dashboard and wizard, so everything on the Status & validation page applies to its results.

climaid ai              # in the terminal
climaid ai --browser    # as a chat page in your web browser

The assistant is a conversational guide, in the terminal or as a chat page in the browser interface (Assistant in the dashboard's menu, or climaid ai --browser). You describe what you want in plain words; it asks for anything missing, shows a plan, runs ClimAID when you say yes, and explains the results.

It is built in and works offline. It does not use an AI language model, local or online: it recognises common requests from keywords and phrases, and every number and sentence about your results comes from ClimAID's own calculations and its rule-based reports. It cannot invent results. When it does not understand something, it says so and suggests what you can type.


An example

you> I want a dengue forecast in Pune for the next 6 months
     I found several districts with that name. Which one?
       1. IND_Pune_MAHARASHTRA  ...
you> 1
     Where is your disease data file? ...
you> data/pune_dengue.xlsx
     Data file looks good: 156 monthly records from January 2012 to December 2024 ...
     Here's the plan: a Dengue forecast for IND_Pune_MAHARASHTRA, 6 months ahead ...
     Shall I run it? (yes / no, or tell me what to change)
you> quick tuning, use data up to December 2023
you> yes
     ... expected cases month by month, likely ranges, trust rating, link to the full report
you> how reliable is this?
you> what about 2050 under high emissions?

In the browser

The chat page works the same way, with a few conveniences:

  • 📎 Upload your disease data (and, outside South Asia, your own climate data or climate projections) instead of typing a path.
  • Suggestion buttons under the conversation: Yes, run it, district choices, follow-up questions.
  • Long runs continue in the background; the page shows how long the run has taken so far. Keep the climaid browse (or climaid ai --browser) terminal open until it finishes.
  • Links open the saved report and the bundled documentation directly.

Each browser keeps its own conversation; Start a new conversation clears it.


What it can do

Ask for Example
A forecast of the coming months forecast dengue in Pune for the next 12 months
A climate-change outlook what could happen by 2050 under SSP2-4.5 and SSP5-8.5?
How much to trust a result how reliable is this?
An explanation of results explain the results, open the report
Answers from the documentation what is WIS?, why is 2020 excluded?, how many trials do I need?
How the methods work methods, how does calibration work?, then more detail
How a run was made how was this forecast made?
Finding a district find district Kathmandu
Your current settings settings

Explaining the methods

Type methods for a numbered list of topics, or ask about any of them directly ("how does tuning work?", "what is the renewal model?", "how are climate models bias-corrected?"). Each answer comes in plain language first; say more detail for the technical version (equations, defaults, lag ranges), with a link to the documentation page.

Forecasting Climate outlook General
Data checks; the COVID-19 period; climate inputs and lags; distributed lags; the models; the baseline; the transmission (renewal) model; residual learning and the ensemble; tuning; backtests; calibrated likely ranges; scoring (WIS) How the outlook is built; bias correction; seasonal vs anomaly response; lag structure; pooling districts; thermal-suitability curve Overview; the trust rating; avoiding leakage; ClimAID v1; limitations

After a run, "how was this forecast made?" describes that particular run from its own record: the data period, the models compared, which models tuning improved and which kept their defaults, the backtest start dates, the calibration, the COVID-19 handling and why the model in the summary was chosen.

Settings it understands anywhere in a sentence:

Setting Say
Data file a path ending in .csv, .xlsx or .xls (in quotes if it has spaces)
Forecast start use data up to December 2023, from 2023-12
Months ahead 6 months, next year (1–36; for longer, use the outlook)
Tuning quick / balanced / thorough, or 50 trials
COVID-19 period keep 2020, exclude 2020-03 to 2020-12
Emissions pathways SSP2-4.5, high emissions, all scenarios; also include ... adds to the list
Outlook end by 2050, to 2080
Your own climate data my climate file is "weather.csv" (places outside the built-in South Asia data); for outlooks also my projection file is "cmip6.csv"

The assistant always shows the plan and waits for yes before running anything. Reports are saved in climaid_outputs/reports; ClimAID's progress messages go to climaid_outputs/assistant_log.txt.


Limits

  • It understands common phrasings, not every possible sentence. Rephrase, or type help for examples.
  • It runs one district at a time with ClimAID's default v2 models. Pooling several districts (extra_districts), choosing individual models and other advanced options are available from Python (see the v2 API) and the dashboard.
  • Documentation answers are short summaries with a link; the linked page has the full explanation.
  • The first run for a region downloads ClimAID's climate data, which needs an internet connection once.