Ask AI

What you can ask

• Data questions - figures, trends, rankings: "Which 5 customers have the highest total order amount in public.orders on Sales DWH?", "Total order amount per month in 2024 as a line chart." The assistant writes and runs a SELECT, and answers with prose, one or two charts, the query and a sample of the rows. A question asking for several breakdowns at once ("sales per region, per state and per store type") gets one query per breakdown, each with its own chart, up to six charts in all. A table named in the question (public.orders) is the one queried, never a model or copy built from it. A follow-up such as "now only 2024" or "split it by region" refines the same queries.

• Metadata questions - what SelectZero knows about the data: "What is the lineage of dwh.orders?", "Which validations cover public.orders and when did they last run?", "Which columns carry PII?". Answered from the catalog, glossary, lineage, execution history and audit, with the evidence shown as clickable tables.

• How-to questions - "How do I schedule a test suite?" Answered from the documentation shipped with the product, with links to the matching pages. No documentation is fetched from the internet.

Ambiguous questions get a question back instead of a guess - for example when the same dataset exists on two connections - with the choices offered as buttons.

Reading an answer

A data answer is prose plus, in order. The prose may set key figures in bold, object and column names as code, and steps or items as lists.

• Charts - bar, stacked bar, line or doughnut, chosen for the shape of the result. Every shortened chart says so underneath, and the full rows stay in the result data:

  • A bar chart with more than 40 categories shows the 40 largest (a time axis shows its first 40 periods).

  • Past four groups, a stacked bar shows the three largest and sums the rest into a grey Other segment; side-by-side bars and lines show the four largest and leave the rest out.

  • A doughnut names its five largest slices and shows the rest as one grey Other slice; its tooltip gives each slice's share in percent.

  • A chart whose result has several rows per bar (months under a quarter axis) is not drawn, rather than showing one row's value as the total.

  • Charts never reuse a colour for two different things: series past the palette are left out with a note rather than coloured alike.

  • A chart with nothing to plot (no row with a label and a numeric value) is left out.

• Data sources - the objects the assistant inspected and used, each opening its catalog detail.

• Then, per query (numbered Query 1 of 3 when there are several):

  • Result data - the rows the query returned, up to 200 (click the label or the arrow to open them). The first 10 are shown, Show all lists the rest, and sorting a column sorts every row. Long decimals are rounded, and a column of shares (values between 0 and 1 that add up to 1) reads as percentages; sorting uses the values as returned.

  • Generated SQL - the exact query, with a copy button. A long query shows its first lines and Show the whole query shows the rest; the copy button always copies all of it. If the query could not be run, the SQL and the reason are still shown, so it can be run by hand.

A metadata answer is a short summary with findings, followed by evidence tables:

• Object lineage and Column lineage sections, one per object looked up, listing the related objects with their direction (upstream or downstream), level (1 = a direct relation) and relation type. The Graph toggle shows the same lineage as the interactive graph from the catalog, inside the chat; Open lineage opens the object's catalog detail on its Lineage tab.

• Validations, Glossary terms and Catalog objects sections for the entities the answer is about, each row opening the entity. A question about validations always gets its Validations table, built from what the assistant actually found, with each validation's last result and last run. Glossary terms show where they are linked (the first three columns, then a count). A question naming a catalog object gets its Catalog objects row. Glossary terms linked in descriptions show as their names.

Evidence tables show their first 10 rows with Show all, like the result data: the conversation is the only thing that scrolls. Lineage tables appear with metadata answers only; a data answer does not list the lineage the assistant looked at on the way.

A how-to answer ends with a Documentation section: the matching pages and sections, opening the online documentation in a new tab.

Conversations

Conversations are listed on the left and survive reloads and logins. New chat starts a fresh thread; Remove from history deletes one. They are stored encrypted, per user, and are deleted with the user. The newest 50 conversations are kept, and the newest 30 turns of each.

The answer to a follow-up is built from the questions asked so far and the query that answered the latest of them, so a follow-up can refer to "it" or "the same period".

While the assistant works

An answer takes several steps - the assistant may search the catalog, inspect columns and run a query before it answers - and the step it is on is shown under the input. Cancel stops after the current step, and nothing is saved for a cancelled question. Answers keep running if the page is closed and are in the conversation when it is opened again.

When there is no answer

• A question back - the ask is ambiguous in a way that changes the answer. Pick one of the offered choices, or reply in your own words.

• No matching data found - the assistant searched and found nothing that answers the question. Naming the business area or the table more explicitly helps, as does enriching the catalog with descriptions so the assistant can recognise the relevant objects.

• The question could not be answered - the assistant ran out of steps, every query it ran failed (for example because the connection could not be reached), the AI model ended without an answer, or the question needs data that you may not query. For a data question the last usually means query execution is not permitted for you on the connection involved (see below); otherwise a more specific question, naming the connection or tables, gets there in fewer steps.

• The query could not be run - the query was written but refused, failed or stopped. The SQL is shown with the reason and a hint that fits it: a query the safety check could not read may well be valid for your database, so copy it and run it yourself; a query stopped at the 60-second limit needs a shorter period or fewer rows; a database error names what to check. A failed query is never run a second time.

Privileges and data access

Ask AI sees exactly what the asking user may see:

• The connections it can search are the ones where the user has View catalog.

• Queries only run on connections where the user has View sample data. Without it on a connection, the assistant still writes the query and shows it, but does not run it. Without it on any connection, data questions are answered with the SQL only, and only over catalog objects the assistant inspected: a question that needs anything else (system tables, storage sizes) is answered as not possible without data access, never with an unchecked query.

• Every query passes a read-only check: a single plain SELECT, no other statement, and at most six query runs per question, each capped at 60 seconds. Anything else is refused before it reaches the database. Row-level access beyond that is the connection's own database user.

• Lineage, catalog and validation lookups respect the same connection and schema limits as the catalog pages.

What is sent to the AI provider depends on the instance settings. Object and column names, and the metadata the tools return, are always sent. Row values from query results are sent only when Include sample data values is on under General settings; with it off, the assistant sees the shape of a result but not its values, while the user still gets the full rows and charts. See What leaves your deployment.

Documentation search

How-to questions are answered from the documentation bundled inside the application, so they work without internet access. Out of the box the search is full-text. An administrator can additionally build a semantic index, which lets a question match a page by meaning rather than by the words it uses:

1. Under Settings - General - AI assistant, set Embedding model if the provider's default is not wanted (Azure OpenAI needs the name of an embedding deployment; Anthropic has no embeddings API, so the index stays unavailable there).

2. Press Build semantic docs index. The documentation is embedded once through the configured provider, and the result is cached on the server.

The status line under the button says whether the index is built, and for which model. Changing the provider, the embedding model or the product version (which changes the documentation) makes the index stale; search then silently falls back to full-text until it is rebuilt. Building it is the only time the documentation is sent to the provider.

Provider limits

Each question is several model calls carrying catalog metadata and query results, which can exhaust a low usage tier's per-minute allowance in one go. The assistant waits out one short cool-down on its own; past that the question fails with a message saying the provider is rate limiting. See Provider rate limits for what to do about it.