Docs
Docs /Prism AI /Introduction
Prism AIIntroduction

Prism AI

An agent inside the IDE that reads your open tab and your schema, dry-runs the SQL it writes, and edits the query in place, with your approval by default. It costs credits from your plan, or runs on a subscription you already have.

right panel
Prism×
Prism AI
Your SQL assistant for BigQuery
What can Prism do?Optimize my queryExplain this queryHelp me write a query
Plan, @ for context, / for commands Approveauto
A fresh conversation. The four buttons are ready-made prompts; the input takes @ mentions and / commands.

Opening it

I (CtrlI on Windows and Linux) toggles the panel, and so does the triangle icon in the top bar. The panel header has New conversation, Conversation history, Settings, and a credit counter such as 87/100.

The header also shows the attached tab — the tab the agent reads from and writes to. It follows the active tab; click it to pin a different one.

  • Signed in, or in the Playground (10 requests a day there)
  • Settings → AI Features → Enable Prism AI Features is on (it is by default)

What it can see

Nothing until you ask. Each message carries the mode, the attached tab, and whatever you attach with @:

MentionWhat goes to the model
@project.dataset.table, @dataset, @projectSchema metadata: columns, types, partitioning. No rows
@graphA property graph’s labels, properties, node and edge tables
@Tab nameThat tab’s SQL and its last job (bytes, cost, duration)
@Tab name under ResultsThe result’s schema and row count; rows only if you allow it

Beyond that the agent uses tools to look things up on its own: it can browse the schema, read tabs, validate SQL with a dry run, and read BigQuery’s documentation; none of which needs confirmation. Anything that scans data or changes a tab is governed by the mode.

What it can do

  • Write a query from a description, or change the one in the attached tab, as a diff you apply or reject.
  • Run SELECT queries and show the result in the chat. INSERT, UPDATE, DELETE, and DDL are rejected before they reach BigQuery.
  • Explain, debug, and optimize: it validates its own SQL, sees the dry-run estimate, and knows about partition and cluster columns.
  • Sample a table’s data at no cost, draw a chart, create tabs and notebook cells, search your saved queries.

Where your data goes

Messages, attached SQL, and schema metadata are sent through Querylab.io’s backend to the provider you chose. The backend relays and doesn’t store the content it relays; the conversation itself is kept in your browser’s storage and on Querylab.io so you can resume it.

Query results are the exception. By default the agent runs a query and sees only its schema and row count. The panel says AI cannot see query results — only the schema is shared and offers Share with AI per result. Settings → AI Features → Share Query Results Preview has three values:

ValueBehaviour
Disabled (never share)Rows never leave the browser
Ask each time (default)A Share Query Results? dialog, with Remember my choice
Always allowRows are shared without asking

The same setting covers table samples. Tool results are capped in size before they go to the model; if a result is too big, rows are dropped, not the query.

On the desktop app you can bypass Querylab.io’s backend entirely and run the agent through a CLI you already have installed, or on your own API key — see CLI integration.

Models and credits

The model dropdown next to the mode selector lists Querylab.io Cloud models and, on desktop, Your Subscriptions. Auto picks a model per message from the complexity of the request and the SQL involved; its cost shows as varies. Every other model shows a multiplier (1x, 7x), which is the credits one message costs. Your plan sets the monthly credit allowance; the counter in the panel header and Settings → AI Features → AI Providers show what’s left and when it resets.

The model is fixed for the length of a conversation. To change it, start a new one.