AI Studio screens reference

Field-level reference for the AI Studio screens in the Explorer: Models, Embeddings, Knowledge Bases, Prompts, RAG Pipelines, AI Agents, AI Workflows and Evaluations. For concepts and walkthroughs see AI Studio concepts; for guardrails, traces and channels see AI operations and safety.

Models

Explorer: AI Studio > Models. Page title AI providers. Tabs: Catalog, My providers, Usage, LLM Providers, Embedding Providers, API keys, Vector Database.

Catalog

A table of every model available to the workspace.

Column	Description
Model	Model code.
Provider	Provider code. Filterable.
Type	Chat, embedding and similar. Filterable.
Capabilities	What the model supports (for example tool calling).
Context	Context window in tokens.
Price / 1M tokens (in / out)	Input and output price.
Source	Platform catalog or your own.

Summary figures: Models, Providers, Embedding models, Your own models.

My providers

Use this to add a model server that is not in the platform catalog. No providers of your own yet. Add one to use a model that is not in the platform catalog. is shown when empty.

Field	Description
Name	For example Groq or Company gateway.
Provider code	Lowercase letters, digits and dashes. Cannot be changed later.
Provider type	OPENAI_COMPATIBLE, AZURE_OPENAI, GOOGLE, AWS_BEDROCK or CUSTOM. Type-specific fields follow (deployment and API version for Azure, AWS region for Bedrock, request template JSON and response text path for Custom).
API key (optional)	Stored encrypted and never shown again. If the server cannot store keys: This server cannot store keys yet; ask your platform administrator.

Adding a model to a provider:

Field	Description
Model name	Exactly as your server expects it, for example the model id.
Type	Model kind.
Context window in tokens (optional)
Vector dimensions (optional)	Embedding models only. Any length works; run Test after saving.

Test calls the model and shows a success or error result.

Usage

Period selector, then totals and a calls-per-day chart: Calls (with failed count), Tokens in, Tokens out, Average latency, Estimated cost. A table breaks usage down with columns Calls, Tokens (in / out), Avg latency, Errors, Cost. No model calls were recorded in this period. Run a prompt or an agent and they appear here. when empty. Endpoint: /api/v1/ai/usage.

Budgets limit spending:

Field	Description
Name
Applies to	Whole organisation, Users, Agents, Applications, or Development workspaces (plugin id).
Which one	* means each separately.
Period	Per day or Per month.
Token limit, Cost limit	Cost is in the currency of your model prices.
When reached	Alert only, or Stop further AI calls until the next period.

Endpoint: /api/v1/ai/budgets.

LLM Providers and model profiles

A model profile is a named choice of model that prompts and agents refer to, so the model can change without editing them.

Field	Description
Profile code
Model
Temperature	0 to 2. Lower is more predictable. Optional.
Max output tokens	Optional.
Fallback profile	Used when this profile's model is unavailable.

A warning is shown when the selected model has no tool calling and the profile is used by an agent that needs tools.

API keys

Keys stored per provider. They are encrypted on save and never shown again.

Vector Database

Status of the vector store used for embeddings . Shows whether it is available for the workspace.

Embeddings

Explorer: AI Studio > Embeddings. Same page as Models, tab Embedding Providers. An embedding model turns text into vectors for meaning search. Choose the model on a knowledge base; changing it makes existing passages need re-indexing (see Knowledge Bases).

Knowledge Bases

Explorer: AI Studio > Knowledge Bases. A knowledge base holds documents split into searchable passages.

List

Summary: Knowledge bases, Documents, Searchable passages, Documents needing attention. Columns: Knowledge base, Documents, Passages, Health (Empty, Ready or an attention count), Default search.

New knowledge base:

Field	Description
Name	For example HR policies.
Code	Lowercase letters, digits and dashes. Cannot be changed later.
Description (optional)

Documents tab

Columns: Document, Status, Passages, Who can see, Added. Actions: upload files (with upload progress), add text (Title, Content), add a web page (Page address, a public https page; private or internal addresses are not allowed), and Split and index again with the current settings.

Field	Description
Roles (comma separated)	Leave empty for everyone. Otherwise agents and pipelines only use this document's passages for people holding one of these roles (or a role under them).

When the embedding model changes, a notice says how many documents need re-indexing. Meaning search skips them until re-indexed; keyword search still finds them.

Connected sources

Feeds that keep the knowledge base in step with an outside source, with sync runs and crawl results.

Pipeline

Field	Description
Name, Description
How to split	By paragraph, By heading, Question and answer.
Passage size, Overlap
Try it on sample text	Paste a paragraph or two to preview how it is split.

Retrieval test

Shows exactly what an agent or workflow would get for a question.

Field	Description
Ask a question
As a person with roles	Optional, comma separated. Empty means everything.
Search mode	Hybrid, Meaning (vector), Keyword.
Number of results
Minimum similarity

Each result shows its passage and ranks (for example Keyword #3). Nothing matched. Try different words, lower the minimum similarity, or add more documents. is shown when empty. Evaluation dataset code saves the question as a case in an evaluation dataset.

Settings

Default search mode, Default number of results, Default minimum similarity.

Prompts

Explorer: AI Studio > Prompts. A prompt is a versioned template that callers run by code. Endpoints: /api/v1/ai/prompts, /api/v1/ai/prompts/{code}, /api/v1/ai/prompts/{code}/run, /api/v1/ai/prompts/assist.

List

Columns: Prompt, Labels, Versions, Updated. New prompt: Name, Code (lowercase letters, digits and dashes; callers use this code), First draft (optional) using {{variable}} for values that change.

Edit and try

Field	Description
System message (optional)
User message	Use {{variable}} placeholders.
Answer format (JSON Schema, optional)	Forces structured output.
Describe the prompt, or what to improve	Asks the assistant to draft or improve the text.

Running a prompt asks for a value per variable and shows the answer, tokens and any error.

Versions

Columns: Version, Labels, Note, Saved by, When. Actions: put a version's text in the editor, compare with the version before (The messages are the same. Only the note or settings differ. when only metadata changed). What changed? (optional) is the note saved with a version. Labels such as prod point a name at a version; callers that ask for a label always get the version it points to.

RAG Pipelines

Explorer: AI Studio > RAG Pipelines. A pipeline answers a question from a knowledge base in fixed stages. Endpoint: /api/v1/ai/rag.

List columns: Pipeline, Labels, Versions, Updated. New pipeline: Name, Code.

Build and try

Try a question runs the pipeline as it is on screen, saved or not, and shows what every step did.

Stage	Setting
Check the question	Guard on the question: Blocked terms (comma separated), When something is found (Flag and continue, Mask personal data, Block the run).
Rewrite the question	Optional rewrite before searching.
Find passages	Search mode, Passages to keep, Minimum similarity.
Re-rank passages	Keep the best.
Build the context	Context size.
Write the answer	Prompt (a library prompt with {{question}} and {{context}}) with Prompt label, or System message and User message written here. Answer when nothing is found.
Check the answer	Guard on the answer, same options as the first guard.

Versions and labels

Columns: Version, Labels, Note, When. What changed? (optional) is the note. Labels such as prod select the version callers get.

AI Agents

Explorer: AI Studio > AI Agents. List columns: Agent ID, Name, Status, Autonomy, Model profile, Tools. Row actions: Execute (test), Publish (draft -> active), Suspend, Delete agent. Describe the agent generates a draft from a description.

Editor tabs

Tab	Content
General	Agent ID, Name.
Purpose & Instructions	Purpose (short description of what the agent is for) and Instructions. Take instructions from the prompt library makes the prompt's system and user messages the instructions; the prompt can use {{intent}}; the prod label is used.
Tools	The tools the agent may call.
Permissions	Permission key rows.
Autonomy	One of the levels below.
Model	Model profile.
Knowledge	Linked knowledge bases. No knowledge bases yet. Create one under Knowledge Bases, then link it here. when none exist.
Guardrails	Per-agent guardrails.
Feedback	Ratings and comments on answers.
Versions	Saved versions and rollback.
Memory & skills	Remembered facts and skills.
Analytics	Usage and outcomes.
Lifecycle	Publish, suspend and history.
Voice	Phone and voice use of the agent. See Bot designer and voice.

Autonomy levels

Level	Behaviour
READ	Read-only. Refuses any tool whose risk level is not LOW, or whose code looks mutating (.create, .submit, .update, .delete, .cancel, .approve).
RECOMMEND	Search and analysis only, with the same refusal of mutating tools.
PREPARE	May create drafts but refuses submit-type actions by name.
EXECUTE_WITH_APPROVAL	Approval-gate integration is not built yet. Every mutating call at this level is currently refused with a reason; none is silently allowed.
AUTONOMOUS	Same limitation: every mutating call is currently refused rather than auto-executed.

AI Workflows

Explorer: AI Studio > AI Workflows. AI steps are nodes inside the normal Workflow Designer; the node types are listed in Node reference and the guide AI workflow nodes. An AI node calls a prompt, a RAG pipeline or an agent by code, so the model, text and knowledge can change without editing the workflow.

Evaluations

Explorer: AI Studio > Evaluations. An evaluation dataset is a set of cases run against a prompt, a RAG pipeline or an agent. Endpoint: /api/v1/ai/evals/datasets, runs under /api/v1/ai/evals/runs/{id}.

List columns: Dataset, Tests (what it tests), Cases, Last run. New dataset: Name (for example Leave policy questions), Code, What it tests (A prompt, A RAG pipeline, An agent).

Cases

Field	Description
Case name (optional)
Question
Values for the prompt's variables	One per line as name: value, for example document: the leave policy.
Expected (what a good answer contains)

Scorers

At least one scorer is required: No scorers yet. Add at least one so answers can be judged.

Scorer	Passes when
Contains	The answer includes the text (the case's expected text unless you set one).
Does not contain	The answer does not include the text.
Equals	The whole answer equals the text, ignoring case and spaces.
Matches pattern	The answer matches a regular expression.
JSON field equals	The answer holds JSON and the named field equals the text, for example risk equals HIGH. Needs Field.
Graded by a model	A model grades the answer from 0 to 10 against your rubric. Pass threshold defaults to 0.7. Costs one more call per case.
Agent calls a tool	Agents only: the run must call this tool (a code, or a prefix ending in a dot, such as entity.leave_request.).
Agent does not call a tool	Agents only: the run must not call this tool.

Runs

Version to test selects the version. Run list: When, Tested, Passed, Avg score, Avg time. A run's results show Case, Result, Vs compared run, Score and Answer, so two runs can be compared.

Related

AI Studio concepts, AI operations and safety, Bot designer and voice, Use AI Studio from your plugin, the CLI, the SDK and MCP.