Playground
What Playground is for
Playground is where you test the current RAG agent against real bucket-backed workspace knowledge. It is not a mock chat surface; it uses the same API-key scope, agent variants, and retrieval policy that integrations use.
Use Playground to answer one question:
Is this agent variant ready to answer from the buckets it is allowed to search?
What you can do
- start a fresh chat
- select the active agent profile or default
calypso-rag-agent - select the active project API key
- pick a response mode (Fast, Medium, or Extended) for the answer
- send free-form prompts
- inspect streaming answers
- inspect grounded source output
- attach files through the bucket-backed agent upload flow when the agent supports it
- move to bucket setup, Agent Builder, or integration docs when the test exposes a gap
Response mode
The response mode control next to the prompt lets you trade speed for answer depth per turn:
- Fast — quickest answers (Gemini Flash Lite)
- Medium — balanced (Gemini Flash)
- Extended — strongest answers at higher latency (Gemini Pro)
Over the API this is the OpenAI-standard reasoning_effort (low / medium / high). The selected mode is capped to your plan's ceiling. See Controlling answer effort for details.
What should be ready first
Playground is much more useful when these are already in place:
- at least one indexed source in a knowledge bucket
- a reviewed default or named agent policy connected to the expected bucket
- a real active project API key
Without those pieces, the UI can still send a prompt, but the result is not a useful signal about grounded retrieval quality.
Bucket-backed file attachments
When you attach a file for an agent conversation, the upload is still bucket-backed.
Playground should:
- auto-select the bucket when the selected agent variant has exactly one active bucket
- ask you to choose a bucket when the variant has multiple active buckets
- block the upload when the variant has no active bucket
Use this path for one-off files that support the current conversation. Use Knowledge Buckets or the Upload API when the source should become reusable governed knowledge.
Settings rail behavior
The settings rail gives you fast access to three important controls for the currently selected agent:
- whether grounded source titles are included
- whether small talk is allowed
- the current role briefing
For supported agents, these controls can persist changes back to the saved agent policy rather than acting as temporary page-local switches. Treat changes here as policy changes for the selected variant.
Grounding states
The most important interpretation work in Playground is not whether the assistant replied. It is how grounding surfaced.
Available
Grounded sources were returned and can be inspected directly.
Hidden
Grounding likely happened, but source titles are intentionally hidden by policy.
Missing
The run appears to expect grounding output, but source information did not surface cleanly. Check the selected key, agent variant, and bucket readiness before treating the answer as valid.
None
No grounded sources were returned.
How to use Playground well
Use this rhythm:
- test one prompt that should succeed cleanly
- test one prompt that should fail safely
- test one prompt that exposes missing source coverage
- test any required file attachment with the selected bucket
- move to an integration surface only after the behavior is understandable