Ask an AI about your schema and your plan — without sending it your data

A general-purpose AI model can read a SELECT statement and tell you what it does. It cannot tell you why the optimizer chose a master map scan over a bitmap index, or whether the table's TuneTable statistics are stale enough to make the whole plan suspect — because that is IRIS knowledge, not SQL knowledge. The AI Assistant, new in 4.02, adds that domain knowledge to the model you already trust, and does it without asking you to send your data anywhere to get the answer.

Three actions, in their own menu

An AI menu sits in the menu bar between Tools and Window, with three entries. All three act on the active SQL editor — open or click into the editor you mean before choosing one; there is no toolbar button and no grid context-menu entry for them in this release.

• Generate SQL from description… turns a plain-language description, plus the current schema context, into a SQL suggestion inserted into the editor as a marked block at the cursor. It is never executed automatically — you review it and run it yourself, the same two-step path as any other statement you write.
• Explain statement explains the statement at the cursor: what it reads, which indexes could serve it, and where it might be slow.
• Explain execution plan explains an actual execution plan in plain language — on IRIS and Caché, specifically calling out master-map scans, temporary files, bitmap versus standard index use, and whether the table's TuneTable statistics look current enough to trust the plan the optimizer chose.

On InterSystems IRIS or Caché all three carry that extra depth — maps, the Master Map, bitmap versus standard indexes, TuneTable state — on top of what a general-purpose model already knows about SQL, so an explanation talks about your database engine, not SQL in the abstract. On a non-IRIS connection the same three actions work with whatever the schema metadata offers, without that extra depth.

What is sent, and what never is

This is the part to understand before turning the AI Assistant on for a connection, not an afterthought at the end of the page.

Never sent, under any setting:

• Row data, table contents, global contents, or any sampled column values.
• Credentials or connection details — password, JDBC URL, host name, user name, connection name. The provider only ever learns that the connection is, generically, "an InterSystems IRIS database" or the equivalent product name.
• Your license or configuration data.

Sent only after you opt in, per connection, off by default:

Literal masking is on by default: a value like WHERE last_name = 'Müller' is rewritten to WHERE last_name = ? before anything is sent, because the literal itself is exactly the row-level data the rule above keeps out. You can switch this off per connection if you need the AI to see literal values — working with test data you are comfortable sharing, for instance — and the send preview, below, always shows the exact text that goes out, masked or not.

Nothing is sent to Data Prudentia. This feature is bring-your-own-API-key: you configure a provider and it talks directly, over your own network connection, to the provider you chose. Data Prudentia is never in that path and never sees the prompt, the schema, or the answer.

Bring your own provider

Configure at least one provider under Tools > Settings… > AI Assistant. Four provider types are offered: ANTHROPIC, OPENAI, AZURE_OPENAI, and CUSTOM — any other OpenAI-compatible endpoint, self-hosted or otherwise. The base URL is optional for Anthropic and OpenAI, which default to the vendor's own endpoint, and mandatory for Azure OpenAI and a custom endpoint, since neither has one to fall back on. Exactly one provider is active at a time; that is the one the three actions use.

The key is stored the same way a connection password is — Windows DPAPI — never in plain text, never exported with a connection file, and never written to the log; the log only records that a key was set.

Your own endpoint. Point a CUSTOM or AZURE_OPENAI provider's base URL at a self-hosted, OpenAI-compatible endpoint — Ollama, vLLM, LM Studio, or an Azure OpenAI deployment in your own tenant — instead of a public cloud API. For an organisation where even schema names must not leave the building, this is the intended path: the prompt never leaves your network, and nothing else about the feature's behaviour changes.

Turned on per connection, not globally

Sending anything beyond your own typed text needs an explicit opt-in on the connection itself — a global on/off switch does not exist, deliberately: a connection to a scratch database and a connection to a patient database sitting side by side in the same Server Navigator need to be trusted independently. Open the connection's Properties… dialogue and switch to the AI tab:

• Allow sending schema metadata and statement text to the configured AI provider — off by default, including after an update and for an imported connection profile.
• Include SQL literal values — off by default, and only selectable once the first checkbox is on.

Without the opt-in, the three actions are not hidden: running one writes an explanation into the result panel instead of silently doing nothing. You still open the tab yourself; the message names the way, it is not a button. The neighbouring case — no provider configured at all — does offer an Open AI Settings… button that takes you straight to the settings page.

The send preview

The first time you use any of the three actions on a connection in a session, a dialogue titled AI Assistant - review what will be sent shows the complete text about to be sent — system instructions, schema/statement/plan context, your request — with OK/Cancel. It is remembered for the rest of the session only: restart SQL DATA LENS and the next request on that connection asks again.

After a request has gone out, the same text is reachable any time from the Show What Was Sent… button on the result panel. Tick Always show the "what will be sent" preview before sending on the AI Assistant settings page to see it before every single request, not just the first per connection — off by default, and useful in regulated environments where a per-request check is part of the process. The preview is not a description of what gets sent; it is the exact string, so a masking mistake or an unexpected piece of context is visible before it goes anywhere.

What you are paying for

Not model usage — that runs on your own API key, billed by the provider you chose, and SQL DATA LENS never sees or marks up that cost. What you pay for is the integration: the schema and plan context builder, the IRIS domain knowledge built into the prompt, the literal masking, and the key storage. Saying that plainly costs us nothing to admit and should earn more trust than pretending otherwise would.

Editions and the trial

The AI Assistant ships in the Named User, Team and Enterprise editions, and in the 30-day trial included with every download — it is not part of the Free Edition. Running one of the three actions there shows a message naming the edition you need rather than hiding the menu entries, the same way the rest of the paid-only surface behaves. See pricing for what each edition includes, or write to sales@sqldatalens.com with questions a trial cannot answer.

Next step

The part worth trying first is the send preview itself — turn the opt-in on for a test connection, run Explain execution plan on a slow query, and read the exact text before you click OK. Download SQL DATA LENS and see what it says about a plan you already know the answer to.