Custom Default Model for Quick Answer Follow-ups

MaCl0wSt

This feature would allow users to specify a default AI model for the 'Continue in Assistant' function when following up from Quick Answer. Currently, clicking this button uses whatever model was last used in the Assistant, which can lead to unintended routing of questions to custom assistants that might not be suitable for the follow-up query. Adding this configuration option would prevent scenarios where questions are sent to non-ideal custom assistants.

The implementation would work as follows:

Add a setting in the AI preferences to specify a default model for Quick Answer follow-ups
When clicking 'Continue in Assistant' from Quick Answer, the system would use this specified default instead of the last-used model
Users could still manually switch models after continuing to Assistant if needed

Example usage scenario:
A user has created custom assistants for specific tasks like code review or data analysis. While using Quick Answer for a general query, they want to explore the topic further by clicking 'Continue in Assistant'. With this feature, even if they were previously using a specialized custom assistant, the follow-up would automatically use their preferred general-purpose model instead of potentially routing to a limited-capability custom assistant.

This extends the existing Quick Answer-to-Assistant workflow by adding a layer of user control over the transition, making the feature more reliable and predictable for users who work with multiple assistant models.

sprite1091

I've noticed the model supporting Quick Answer changing opaquely from Mistral Small to Gemini Flash 2.0, the latter providing generally shorter, less-informative answers to my queries. I'd signed up for an annual plan in part because of the usefulness of Quick Answer, which in my opinion, is less consistently useful, now.

I'd love for the model for Quick Answer to be user-configurable, even if the choice of model is limited to less-expensive, low-latency LLMs.

I'd envision the introduction of a new account setting for Search, with a drop-down to choose, from a limited selection, the model from which Quick Answers are supplied. In addition, the Quick Answer Tool Tip can display the model that was used.

Luis

sprite1091 We're currently working on Quick Answer and plan to include information about the base model in our documentation.

We consider Quick Answer a core search feature, and our responsibility to ensure we select the best possible setup (including the base model) for its intended purpose.

providing generally shorter, less-informative answers to my queries.

You mentioned that Quick Answer is providing generally shorter, less-informative answers to your queries. This kind of performance feedback is very important to us. Is this still your experience? If so, could you provide a few examples?

sprite1091

Luis Hey Luis, my comment was more to express a non-specific preference for a kind of LLM response and level of verbosity and information, not so much concerning a particular query's QA; and because I cannot change the model myself, A-B testing to showcase the difference here isn't possible.

That said, the only reason I'd noticed the swap from Mistral to Flash 2.0 was due to the decline in quality and results-synthesized information, but this is very much a subjective matter of taste, and so allowing users to configure the model which best suits them would be awesome.

Thibaultmol

@Luis now that the Quick Assistant has launched, I guess this problem is kind of solved right?
That is the model that is selected for follow-up questions from quick answer