Playground
Build › Playground is where you try a model before you build on it. You chat with any chat model your tenant can call, as yourself or as an application holding a given API key, and optionally ground the answers in one collection. Every turn goes through the gateway: the tenant’s grants, budgets and guardrails apply, and the turn is metered.
To try an agent with its tools, skills and collaborators, use Preview and test; to try a published assistant, use Preview chat in Assistants.
Who can open it
Section titled “Who can open it”| Role | What they can do |
|---|---|
| Admin | Chat, paste a key, Generate key |
| Builder | Chat, paste a key |
| Auditor | No access |
| Reviewer | No access |
| End user | No access |
| Platform admin | Holds every capability. Chooses the tenant to chat in, since a platform admin has no tenant of their own to meter a turn to |
What you see
Section titled “What you see”- The header, Playground, and one line: chat with a deployed model as yourself; paste an API key to see what an application with it gets.
- The controls, in a row:
- Tenant, when you have more than one.
- Model: the chat models the tenant can call right now. Selected by default is the first.
- Ground in collection: None (model only), or one collection, shown as database / collection.
- API key (optional): leave empty to chat as yourself.
- Generate key and a link to Keys and budgets, for Admins.
- The chat, headed Chat, with Restart conversation in its header. Its start screen says “Chat with your deployed model through the metered gateway.” and offers two starters: What can you help me with? and Write a haiku about the cloud.
The Model list holds only models that can answer a chat now: granted to the tenant, deployed and ready, or served by an enabled cloud provider. Embedding models and rerankers are not in it.
Empty and error states
Section titled “Empty and error states”| You see | What it means |
|---|---|
| No chat model is deployed for this tenant with Go to Models | The tenant has no chat model it can call. An embedding model cannot chat. Deploy a chat model, or connect an external one, then come back |
| The deployed models could not be loaded with Try again | The list could not be read. It is not the same as no models |
| The model did not answer. and a reason | The turn was refused or failed: the model is not granted, the budget is spent, the key is wrong or expired, or the model is unavailable |
| The key was not generated | Generate key failed. When the reason is “This tenant has no project to hold a key. A platform admin creates one.”, ask a platform admin for a project |
What you can do
Section titled “What you can do”Chat with a model
Section titled “Chat with a model”- Choose the Model.
- Leave Ground in collection at None (model only).
- Type in Type something… and press Enter, or choose a starter.
The answer streams in. Under it, a line gives the token count: Tokens: n in, n out, n in all. Choose Stop response to stop an answer mid-way. Each question carries the conversation so far, up to the last 16 messages.
Ground the answers in a collection
Section titled “Ground the answers in a collection”- In Ground in collection, choose a collection.
- Ask a question.
Olympusx AI retrieves the passages that best match your question from that collection, up to 6, and gives them to the model with the question. The answer ends with Sources: and the numbered documents it drew from. To compare, ask the same question with None (model only).
The collection list holds every collection in every vector database of the tenant. See Collections.
See what an application gets
Section titled “See what an application gets”Paste an API key to chat as the application that holds it. The key’s own model allow-list, budget and expiry apply, and the turn is metered to the key’s project. The key must belong to the tenant you are working in.
- Paste the key in API key (optional).
- Ask a question.
A model the key may not call is refused with the reason.
Generate a key here
Section titled “Generate a key here”Admins can make a key without leaving the page.
- Choose Generate key.
- The new key fills API key (optional).
The key is created in the tenant’s first project, named playground- followed by the date and time, and expires after 90 days. It is a real key: manage or revoke it in Keys and budgets. Its secret is shown only in this field, so copy it now if you need it elsewhere.
Start again
Section titled “Start again”Choose Restart conversation in the chat header. The model forgets the conversation so far. Each turn is metered, and traced when tracing is on.
Limits
Section titled “Limits”| What | Limit |
|---|---|
| Conversation context | The last 16 messages are sent with each question |
| Grounding | One collection at a time; up to 6 passages per question |
| Generated key | Expires after 90 days; created in the tenant’s first project |
| Tenant without a project | Cannot chat as yourself or generate a key until a platform admin creates a project |