Reasoning and quality
Two sets of controls shape how good your answers are: the model tier and reasoning you pick per agent, and the optional conversation analysis that grades finished chats after the fact.
Base and Premium
On most plans you choose a quality tier rather than a raw model:
- Base — fast and economical, the right default for most assistants.
- Premium — a stronger model for harder questions. A Premium reply draws more from your monthly message-credit allowance than a Base reply.
Each tier in the Quality select shows the name of the model it currently uses, for example "Premium · Mistral Medium 3.5". The platform maps the tiers to models and may move a tier to a newer model.
On the BYOK and Custom plans you instead pick the exact model and sampling settings yourself. You set the tier per agent on the agent form — see Agents.
Extended thinking
Some models can "think" through a problem before answering. On Base and Premium plans the platform decides whether a tier thinks, and the Quality select marks it "extended thinking". On BYOK and Custom you switch reasoning on per agent with Enable reasoning. Some models always think (for example Magistral and GLM). For them too, the switch (or, on Base and Premium, the tier) decides whether that thinking counts as on and can be shown as Thinking.
You also decide how much of the assistant's work visitors see, with Show the assistant's work on the Configuration tab. Subtle shows one quiet line with the current step. Tool steps lists every step with its duration and search terms. Thinking adds the model's reasoning, the code it ran and its output; you can pick it only when the agent thinks (on Base and Premium plans, when the selected Quality tier has extended thinking; on BYOK and Custom, when the model can think and reasoning is on). A "Still working" line appears in every mode after 3 seconds without new text.
You can review the steps of each reply in the conversation detail under "steps", including the address and title of up to ten results per web search. In private mode only the type and duration of each step are stored.
Conversation analysis
Turning on Analyze conversations for an agent has the platform review finished chats in the background and grade them:
- a helpfulness verdict — helpful, partial, or unhelpful;
- a knowledge gap flag with a short summary, when the assistant couldn't answer because the material wasn't in its library.
These signals surface in the Quality section of the dashboard, where a knowledge gap is your cue to add the missing answer as a library source. Each review draws a small amount from your monthly allowance, shown on the agent form when you switch it on.
See Conversation analysis for the Quality section in full, and Libraries and sources to close a gap.
See also: Agents and Conversation analysis.