Before you start
You need at least one registered model, and at least one evaluator if you want to score responses.Chat with a model
1
Pick a model
Open Settings with the gear icon. On the Configuration tab, choose a Model. A model card’s Chat with Model button opens Chat with that model already selected.
2
Adjust the generation settings
Still on Configuration:
- Memory – the Use conversation history toggle. When on, the model sees all previous messages. When off, each message is independent.
- Temperature – a slider from 0 to 1 that controls randomness. Lower is more deterministic.
- Max Tokens – a slider from 128 to 4096 that caps the response length.
- System Prompt – an instruction that guides the model’s behavior, such as You are a helpful assistant.
3
Choose evaluators
On the Evaluators tab, check the evaluators you want available for scoring. Search narrows the list, Select all and Clear work in bulk, and a chip counts your selection. Skip this if you only want to chat.
4
Send a message
Type your message and send it. The response streams in as it is generated. The + button starts a new chat.Models that emit chain-of-thought show a Reasoning chip above the answer. It pulses while the model is still thinking, and clicking it expands or collapses the reasoning.
Evaluate a response
1
Evaluate
Point to an assistant message and click Evaluate.
2
Read the results
Evaluation Results runs your selected evaluators against that response. Each row shows the evaluator, a pass or fail mark, its score as a gauge, and a chip per judge model with that judge’s score. Show only concerning metrics narrows the list to low or failing scores, and expanding a row shows its detail.
3
Go deeper or close
Open in Comparison Workspace carries the response into a comparison. Close returns to the chat.