FAQ
Short answers to common questions. Follow the links for detail.
Do I need to be a data scientist to use Pradra Studio?
No. The platform is organized around the business problem, and sensible defaults are provided at each step — including a default algorithm for training. Specialists get full control when they want it, but you can take a capability live from the UI without hand-tuning anything. Start with the Quickstart.
Do I need an SDK?
No. The Studio UI covers the whole lifecycle — data, training, deployment, monitoring. You only reach for an SDK (or plain HTTP) when your application calls a deployed capability. See SDKs & API clients.
What's the difference between a Solution and a Capability?
A Solution is the business problem; a Capability is one specific prediction inside it. One solution usually contains several capabilities that share data and operations. See Solutions and Capabilities.
How does the platform stop a bad model from reaching production?
Every candidate must pass the capability's acceptance criteria on held-out data before it can be promoted — and if any rule fails, promotion is blocked even if the headline metric improved. This applies to human promotions, automated retraining, prompt activations, and fine-tunes alike. There is no side door.
Something went live and it's worse. What now?
Roll back: point production at the previous model (or prompt) version. It is a pointer switch — instant, reversible, no retraining. See Models & evaluation.
Can I use large language models here, or is it classic ML only?
Both, on the same rails. AI Studio builds generative capabilities (generation, summarization, extraction, embeddings), grounded in your documents via knowledge sources, evaluated against golden sets, and served through the same endpoint shape as classic models.
How do I keep generative costs under control?
Every generative call is metered (tokens and USD), and monthly budgets at the organization or capability level are a hard stop: at 80% you're warned, at 100% further calls are refused until the month rolls over or the limit is raised.
Can an AI agent act on its own?
Within hard limits, yes — but anything consequential (an action that changes the real world) always parks for human approval, with no way to configure that away. Every step of every run is traced. See Agents and Agent tools & approvals.
Is my data isolated from other tenants?
Structurally. Data is scoped to your organization and its workspaces; a query from one workspace can never reach another's data, and cross-tenant references simply resolve as "not found". See Organizations & workspaces.
Where do I put database passwords and provider API keys?
In Secrets: centrally stored, write-only, rotated in one place, and impossible to delete while still in use.
Why does my drift page say "skipped" instead of a score?
Because there was too little traffic in the window to judge honestly. The platform reports "not enough evidence" rather than a reassuring zero — a quiet endpoint is flagged as quiet, not healthy. See Drift & retraining.
I hit "plan limit exceeded". Did I lose anything?
No. Plan limits cap creation only — the request to add one more was refused, everything existing is untouched. See your usage under Subscription & plans.