AI Studio
AI Studio is where you build capabilities powered by large language models — summarize a claim, extract fields from an email, draft a customer response — on the same rails as everything else in Pradra Studio: inside a solution, behind the same endpoint shape, gated by the same quality checks.
Generative capability types
| Type | Produces | Typical use |
|---|---|---|
text_generation | Free-form text | Drafting replies, product copy |
summarization | A faithful summary | Case notes, call transcripts |
extraction | Structured fields validated against your contract | Invoices, forms, emails |
embedding | A numeric representation | Search, retrieval, similarity |
Extraction deserves a highlight: its output is checked against the capability's output contract. A field the model failed to find comes back as an explicit null — never as an invented value passed off as data.
Models are pinned, not floating
A generative capability serves through a model adapter — a registered connection to a provider (or your own self-hosted model). When you bind an adapter, the exact model identifier is pinned to the capability version. The provider releasing a newer model changes nothing in production until you decide to move: no silent upgrades, no surprise behavior changes. Provider API keys are stored sealed, or centrally in Secrets.
Working in the Studio
- Go to AI Studio in the sidebar. The list shows your generative capabilities with their status.
- Create walks you through a wizard: pick the solution, the type, the contract, and the model.
- A generative capability's detail page has tabs for each concern:
- Prompts — versioned instructions; see Prompts.
- Model — the pinned adapter and identifier.
- Evaluation — golden sets and quality scores; see Generative evaluation.
- Retrieval — ground answers in your documents; see Knowledge sources.
- Playground — try inputs and see the live output before anything ships.
- Deploying and calling it works exactly like a classic capability — same endpoint shape, same prediction history (plus token/cost columns), same monitoring.
The safety rails
- Prompt injection separation. User input only ever fills the declared placeholders in your prompt — it cannot rewrite the instructions.
- Content filtering. Deterministic filters you configure run on the way in (blocking before the model is called) and on the way out (withholding a flagged response with an explicit notice). Filtered events are recorded and counted.
- The activation gate. A prompt version that has never been evaluated — or that scored below your criteria — cannot become the serving version. See Generative evaluation.
- Budgets. Every call is metered; a blown monthly budget stops the spending. See Budgets & costs.
Next
- Write and version the instructions: Prompts.