Features
Prism AI answers from what you give it and what it looks up itself. You attach tabs, results, and tables with @, the agent calls tools to browse the schema, dry-run SQL, and run queries, and it renders results, charts, and follow-ups back into the panel so you can act on them without leaving the chat.
search_schema · columns fact_mot_test
validate_sql · valid, 1.2 GB, $0.006
execute_query · 20 rows, 1.2 GB
Ford leads by count, but by rate it’s Vauxhall at 31% failed. The query groups on make and divides failed by total tests — I added a filter on completed_date so it only scans this year’s partitions.
Context with @
Type @ in the input (or click the @ button) and pick from open tabs, results, and the schema tree. Up to ten matches show as you type.
| Mention | Attached |
|---|---|
| Tab | The SQL and the last job’s bytes, cost, duration |
| Results | Schema and row count; rows only under the share setting |
| Table | Columns, types, descriptions, partition and cluster keys |
| Dataset, project | Its tables, so the agent can pick |
| Property graph | Labels, properties, node and edge tables — needed before writing GQL |
Attached items appear as pills above the input; the cross removes one. You can also attach files with the paperclip (.sql, .bqsql, .csv, .tsv, .txt, .md, .json, .jsonl, .yaml, .yml, 32 KB in total); their text is inlined into the message.
Commands with /
| Command | Effect |
|---|---|
/clear | Start over: clears the conversation |
/compact | Replaces the history with a summary so a long conversation keeps fitting; add a note after it to say what to keep |
Tools
The agent decides which tools to call; you see each call as a line in the reply. Reading tools never ask; the two that scan data or change a tab follow the mode.
Reading
| Tool | What it does |
|---|---|
search_schema | Browses projects, datasets, tables, columns, and property graphs. Supports * wildcards and multi-word matches |
get_tab, list_tabs | Reads a tab’s SQL and job stats; lists every open tab |
validate_sql | Prism diagnostics plus a BigQuery dry run: validity, cost, referenced tables. Accepts pipe syntax and GQL |
sample_table | Per-column profiles and diverse sample rows through tabledata.list, which is free. Tables only, not views |
get_optimization_tips | Structural cost findings: missing partition filters, SELECT *, join order, clustering |
get_bq_docs | Searches and loads BigQuery’s SQL reference so it doesn’t guess function signatures |
get_skill | Loads a longer playbook (optimize_query, error_resolution, notebooks) before a complex task |
search_saved_queries, get_saved_query | Finds and reads your saved queries |
Changing things
| Tool | What it does |
|---|---|
execute_query | Runs a SELECT, up to 1000 rows. Anything else is rejected |
edit_sql, revert_edit | Replaces or patches the SQL in the attached tab or a notebook cell; one level of undo |
create_tab, rename_tab | New query or notebook tab; descriptive names after a run |
create_notebook_cell, execute_notebook_cell, list_notebook_cells, get_notebook_cell, get_notebook_cell_schema | Multi-cell pipelines in notebooks |
Showing you things
| Tool | Renders as |
|---|---|
update_todo | A progress list at the top of the reply for multi-step work |
create_chart | A chart from result rows, with a Download chart as PNG button |
suggest_follow_ups | One to three chips under the reply; click one to send it |
In the reply
- Query results appear as a grid in the chat with Open in Results Panel and Open in Tab. What the model saw is separate from what you see; the privacy note under the result says which.
- With models that expose it, a collapsed Reasoning block precedes the answer.
- Thumbs up or down on any reply gives feedback; the down-vote asks what went wrong.
- After 25 tool calls in one request the agent pauses and offers Continue, which counts as one more request.
Turning it off or tuning it
The master switch, AI ghost completions, auto-renaming of tabs, the results-sharing policy, and the provider list live in Settings → AI Features → Prism AI; see Settings.