AI and web functions (paid)
Call a model or the network from inside a formula — generate or classify row by row, crawl a page, search, read a file, hit an API. One call per row; the result is an ordinary formula column. Paid feature, billed by your provider.
What they are
An ordinary function turns one column into another — upper, sumif. AI and web functions work the same way; they just swap "compute" for "ask a model" or "fetch from the network":
| Function | What it does | Example |
|---|---|---|
ai.generate |
Hands each row to the model with your prompt | Tag each piece of feedback positive / negative; normalise addresses; pull a model number out of a description |
web.crawl |
Fetches a page and takes the body text | One product URL per row → title and price |
google.search / perplexity.search |
Searches on a keyword, returns a summary | Look up a website and a one-liner for each company name |
twitter.search / reddit.search / wiki.search |
Searches one specific platform | Discussion or encyclopedia entries per topic |
file.extract |
Extracts text from files in an attachment column | A column of PDF contracts → the two party names |
http.request |
Calls any HTTP endpoint | Look up delivery status by the tracking number in each row |
You write them like any other function — on a column, executed per row:
ai.generate("Answer only positive or negative. Is this review positive or negative: ", AA5)The authority on arguments is the function help in the formula editor; see Function reference · AI and web.
Paid, and what it costs
- These belong to the paid AI plugins group; a free license gets the upgrade prompt.
- One model or network call per row. Running
ai.generateover a 5,000-row table is 5,000 calls — try it on a split-off table of a few dozen rows first, confirm the prompt is right, then run it on everything. - Web functions reach the addresses you write; model calls go through the key you configured.
When to use these, and when to use Ask AI
- Same operation on every row, and on new rows as they arrive: use a function.
- Build a table from nothing, or organise a handful of files once: use Ask AI.
Things to watch
- Results vary with the model; they are not strictly reproducible. For closed-set tasks like labelling, list every allowed value in the prompt and set the result column's field type to single-select as a constraint.
- Web content sometimes cannot be fetched (anti-scraping, a login wall). When a result comes back empty, open the address in a browser first.
- What goes out is the content of that row. Keep data that must not leave out of these functions — see Privacy.