An AI SQL generator that runs on your own warehouse.

What "generator" usually means, and what's missing

Most tools that call themselves an AI SQL generator are a text box: you paste a rough description of your tables, and it writes SQL against that description. The catch is obvious once you use it against anything real. You either paste your entire schema by hand every time, or the model quietly guesses at column names and gets some of them wrong. Either way, the SQL it hands back isn't guaranteed to run.

A worked example against BigQuery

Say you have a BigQuery connection with a dataset called analytics_prod, and you type this into the Ask panel:

"Which plan had the most new subscriptions in the last 7 days?"

Because the Ask panel already has schema access through the connection, it doesn't ask you to describe the table. It comes back with:

SELECT plan_name, COUNT(*) AS new_subscriptions
FROM `analytics_prod.subscriptions`
WHERE created_at >= TIMESTAMP_SUB(CURRENT_TIMESTAMP(), INTERVAL 7 DAY)
GROUP BY plan_name
ORDER BY new_subscriptions DESC
LIMIT 1;

Backtick-quoted project.dataset.table, GoogleSQL's TIMESTAMP_SUB, real column names from your actual schema. Point the same question at a Databricks connection instead, and it comes back written against catalog.schema.table with Databricks SQL date functions, no extra instruction needed, because the panel already knows which warehouse it's talking to.

What it's built on: schema context, not a guess

The Ask panel reads table names, column names, and types through the connection you're already using, plus recent query results and errors when you attach them. That's the difference between a generic SQL generator and one that produces something you can run immediately. It's still worth reading the SQL before you run it against anything that writes data, the panel is good, not infallible, and a wrong WHERE clause on a SELECT is a lot cheaper than one on a DELETE.

One sentence, then the SQL

Answers default to a single sentence plus the SQL that produced it, so you can skim it and move on. Type /verbose if you want to see the full reasoning and the step cards showing which tables and columns it actually looked at, useful the first time you're checking whether it read the right dataset.

From generated SQL to a running check

Once the panel writes a query, three buttons sit under it: Insert into editor, Run, and Watch this. That last one is the difference between generating a one-off query and turning it into something that checks itself, click it and the same SQL becomes a Watch with a condition and a destination, no retyping.

What it doesn't do

It doesn't design your schema, tune indexes, or replace understanding what a query does before you run it. It's also not free-floating: it needs your own API key under Settings → AI Assistant and a real connection to read schema from. If you want SQL written against a schema you haven't connected anything to yet, a generic AI chat tool is still the faster path for that narrower case.

Picking a model for the job

The Ask panel doesn't lock you into one vendor. Add a key for Anthropic, OpenAI, Google, or xAI and pick Claude Sonnet 5, Haiku 4.5, GPT-6 Sol, GPT-6 Luna, Gemini 3.5 Flash, Flash-Lite, or Grok 4.20 from the model switcher, or leave it on Auto and let a cheap, fast model handle exploring your schema and recent results while a stronger model writes the final query. That split matters for a generator specifically, most of the cost in generating SQL against a real schema is in reading through table and column names, not in the writing itself, so routing that exploration step to a cheaper model keeps a session of ten or twenty questions from adding up to real money.

From a generated query to a repeatable check

Generating a query is often the first step, not the last one. Once the panel writes something you'd want to see again, checking whether a table stayed empty, whether a count crossed a line, the natural next move isn't to keep asking the same question every morning, it's to turn the SQL into a Watch once and stop asking. That's the same button, Watch this, sitting under any answer the panel gives you, generated SQL or hand-written.