Playground

To chat with a pipeline, go to Playground → Chat. A new chat starts on the project's first live pipeline; pick another from the pipeline menu at the top before you send. Type your question and press Enter, or click one of the example questions. Shift Enter starts a new line, and a message can be up to 8,000 characters.

The answer streams in under the pipeline's name. To stop it early, press Esc or choose the stop button; Grey keeps what it wrote so far. Once you send the first message, the chat stays on that pipeline: choose New chat to switch.

Chat runs the pipeline's saved settings, including changes that aren't published yet, so you can test a pipeline before you publish it. Every built-in role can use the Playground.

If the project has no pipelines, Chat offers Go to pipelines. If the pipeline has no knowledge bases, it offers Attach a knowledge base. See Pipelines and model providers.

Knowledge base only, beside the message box, decides what the model can answer from. It's on by default, and Grey remembers your choice in this browser.

• On: answers use only passages from the pipeline's knowledge bases. When nothing relevant is found, the answer says so and is marked Not found in the knowledge base.
• Off: the model can add general knowledge where the knowledge bases fall short, and says when it does. Those answers are marked General knowledge allowed, or Answered from general knowledge when no passage was found.

If the project's monthly soft cap is set to Pause ingestion, keep retrieval and spending reaches it, answers are replaced by the most relevant passages under the note Monthly budget reached. See Monthly soft cap.

Answers cite their sources with numbered markers, and a Sources row under the answer lists them. Click a marker or a source to see its title, page or timestamp, and the passage it came from. Open source opens that source's details in Sources, and audio and video start at the quoted moment.

The answer's header shows how many passages were sent to the model and roughly how many tokens they used, like 6 passages · ~2,100 tokens. Click it to see each passage as the model received it, with its retrieval score and whether the answer cited it, and a summary such as how many of the top matches fit the token budget. Copy copies the whole context.

The header also shows how long the question's embedding and the model's reply took. Hover over them for the model, token counts and time to the first token.

Hover over an answer to see its actions:

• Copy: copies the answer.
• Ask again: sends the same question again. Only on the latest answer.
• Helpful and Not helpful: rate the answer. Not helpful asks What was wrong?, with an optional note, then Send feedback. Choose a rating again to clear it.
• Mark unanswered: flags a question the pipeline couldn't answer.

Ratings and unanswered questions go to Quality → Feedback for review. See Evaluations and feedback. Your own messages have Copy and Edit in composer, which puts the text back in the message box.

The Run settings button at the top of Chat opens a panel showing the pipeline's status (Live or Draft) and configuration: its model, knowledge bases, search type and number of passages, minimum similarity, rerank, graph, citations and maximum tokens. Edit pipeline opens it in Pipelines.

Under Session overrides, try different values without changing the pipeline:

• Temperature: from 0 to 2, in steps of 0.1.
• Passages (top K): from 1 to 20.

Overrides apply to this browser tab only and are never saved to the pipeline. The reset button beside Session overrides goes back to the pipeline's values.

The list beside Chat holds your own Playground chats, grouped into Today, Yesterday, Previous 7 days, Previous 30 days and Older. A chat is named after its first question. Search them with Search chats, and choose New chat to start another.

Each chat's ⋯ menu has Rename and Delete. Delete asks you to confirm, then removes the chat and its messages from your history. Its request logs and usage stay.

Playground → Conversations lists every conversation in the project: Playground chats from everyone on the team, plus visitors' conversations with your widgets. Calls to the API aren't saved as conversations; they appear in Request Logs. Each row shows the title, pipeline, channel, number of messages, feedback (helpful, not helpful and unanswered counts), who started it and when it was last active.

Search by title, and filter by channel and by pipeline. Click a conversation to read its transcript with citations and the feedback it received. The transcript is read-only.

A note under an answer explains when it didn't finish:

• Cut off at the pipeline's Max tokens limit, or no answer because the model used its whole output limit: raise Max tokens on the pipeline. Open pipeline takes you there.
• The reply stopped early because of an error: the model provider returned an error part-way. Try again with Ask again.
• Stopped before it finished: you stopped it.

If nothing comes back at all, Grey puts your question back in the message box so you can send it again.