AI Features

With an AI provider configured, Veodyn gains a set of assisted flows. Three rules hold everywhere:

1. AI off means absent, not broken. With ai.enabled false, no AI control renders anywhere, and every flow it assists has a manual path.
2. The model writes words; code assigns IDs. Grounding is assembled server-side from what your instance actually has, and a model's pick is looked up against that same list. A suggestion can never link to a query that does not exist or mis-cite a source.
3. Nothing runs or saves without you. Generated SQL lands in the editor for you to read; proposals become objects only when you press Create.

Create with AI

A Create with AI button sits on the Queries, Dashboards and Query Snippets lists (and in the query editor), and on the KPIs and Reports lists where those enterprise features are installed. It opens a chat scoped to that object kind: describe what you want, answer the assistant's follow-ups (replies can carry clickable suggested answers), and when the plan is concrete a proposal card appears in the transcript.

Proposal cards are editable before creation, and each kind shows its substance up front:

• Query: name, description, the generated SQL in full, data source, and a visualization.
• Dashboard: name and the panel list. Missing queries the AI intends to write are disclosed before creation.
• KPI: name, value column, target, unit, direction, cadence.
• Report: an outline with per-section Include checkboxes, each section marked as reusing an existing query or writing a new one.
• Snippet: trigger, description, and the fragment.

Each kind opens with its own question rather than a blank prompt, which is also the clearest statement of what that kind can do:

Kind	Opens with
Query	Tell me what you want to find out and I will work out which table answers it.
Dashboard	Tell me what this dashboard is for. I assemble it from queries that already exist.
KPI	Tell me which number you want to track and what good looks like.
Report	Tell me what this report should explain and who reads it.

The dashboard wording is worth reading closely: it assembles from queries that already exist, so it is not a way to get new SQL written. That opening line is fixed copy, not model output, so it is the same every time and cannot drift.

Nothing happens by accident

Conversations are capped at 12 user turns, enforced by the server as well as the interface, so the cap cannot be talked around.

Nothing is saved as a side effect of talking. An object exists only once you apply its proposal, closing with an unapplied proposal asks whether to discard it rather than silently dropping the work, and reaching the cap or a provider that will not answer offers the manual path to build the thing yourself.

SQL generation in the editor

The prompt bar above the SQL editor turns a description into draft SQL, placed in the editor with the model's rationale shown underneath, never auto-executed. After the first draft the bar becomes "Edit with prompt" for iterating. If you hand-edit the buffer while a generation is in flight, your edit wins and the stale draft is discarded.

Where generation cannot work, the bar says why instead of vanishing: a JSON-based data source takes no SQL, a schema entry that is documentation rather than a table cannot be queried, and a provider outage reads "AI is unavailable right now. Write SQL manually below."

SQL safety

Generated SQL passes the same guardrails as the no-code Visual builder: single statement, SELECT/WITH only, no state-changing keywords, identifiers checked against the live schema, values emitted as escaped literals, and the tables read must belong to the dataset you asked about. A refused generation is retried once with the reason, then handed back to you.

Edit with AI (dashboards)

In a dashboard's edit mode, Edit with AI opens the same chat pointed at the existing dashboard. The proposal shows the change as a diff: which panels are being added, removed, or redrawn, so you approve exactly what will happen.

The digest (enterprise)

An enterprise feature

The digest and the annotation suggestions below are part of the enterprise edition. Their endpoints are absent from a community build's AI provider, and the Home section and the dialog button that call them are absent with them. Everything above this line is community.

The Home page's AI digest is a short list of generated claims about your data, each citing the query or KPI it reads via a badge that links straight to it. Candidates arrive carrying their own source IDs and the model only writes the sentence, so a claim that cannot cite a real source is dropped rather than shown. The digest reads cached results only; it never launches query runs on its own.

Annotation suggestions (enterprise)

In the dashboard annotations dialog, Suggest annotations finds outliers in the actual series and proposes labeled drafts with their time windows. Each is accepted individually and passes the same validation as a manual annotation; a draft that fails is refused with the reason.

Chart authoring

When the AI proposes a chart (in any flow), it also authors the renderer options: column mapping, stacking, axis scale. Those options are sanitized against the same per-type allowlist used for published reports, and core visualization types can infer a sensible column mapping when one is missing, so an AI-authored chart draws and states which columns it is of.

Privacy and authorization posture

• The browser never talks to a model. Every AI request goes through the frontend server, which strips your session cookie before relaying to the provider with a shared instance key.
• Grounding (the catalog of queries offered to the model) is assembled server-side, so a modified client cannot make the model write SQL against tables that do not exist.
• Grounding runs as the instance's service account, so a suggestion can name a query you cannot open; it can never show its contents, because reading results still happens under your own credential.