Customize LLM Parameters (Temperature, Top P, etc.) for Custom Agents | Voters

Customize LLM Parameters (Temperature, Top P, etc.) for Custom Agents

rvkproject

The key parameters users would like to be able to configure are:
1) Temperature: Adjusting the temperature would allow users to control the "creativity" and randomness of the agent's responses, from more deterministic and logical (low temperature) to more varied and exploratory (high temperature).
2) Top P: For OpenAI models, the ability to set the Top P (nucleus sampling) parameter would give users control over the diversity of tokens the model considers, from a broader (high Top P) to a more focused (low Top P) distribution.
3) Frequency Penalty and Presence Penalty: Similarly for OpenAI models, allowing users to adjust the frequency and presence penalties would enable them to encourage or discourage the model from repeating certain words or phrases, or introducing new ones.
4) Response Mode: For Anthropic models, users would like the option to choose between "Full Responses" (more verbose, with explanations) or "Concise" mode (shorter, more direct responses).
Providing these customization options would give greater control over the agent's personality and outputs.
This level of configurability would significantly improve the flexibility and utility of Custom Agents.

November 7, 2024

Brian Sparker

Merged in a post:

[Create Agent] Give us real control: style options, in-painting, model parameters & true multimodal loops

Julien

Hi,

The “Create” agent is great for one-shot image generation, but it feels like a black-box diffusion model with very limited control.

Here are IMO the main key issues:

Over-polished, homogeneous output

• We can’t enable / disable specific style embeddings or presets—results often look too “glossy” for professional use.

No in-painting or targeted edits

• Asking the agent to tweak part of an image or keep certain elements is ignored; it always produces a totally new picture. The chat UI suggests an iterative workflow, which is misleading: in practice it’s just re-prompting a single model.

Hidden diffusion parameters

• Temperature, guidance scale, steps, seed, aspect ratio, etc. are locked away. Power-users need at least an “advanced” panel (or API params) to fine-tune output.

Under-utilised multimodal potential

• You.com offers strong multimodal LLMs, yet “Create” doesn’t loop back through them. We’d love recursive workflows: generate → analyse → refine → regenerate, all inside chat.

There are several impacts to that situation:

• Creative workflows hit a ceiling—can’t iterate or produce brand-consistent assets.

• Extra tools (Photoshop, Stable Diffusion UI) are required, breaking the You.com experience.

• Difficult to meet client briefs when minor edits mean regenerating from scratch.

Here are few of m'y suggestions:

Style / embedding toggles

• Let us choose or exclude style packs (e.g., “photorealistic”, “toon”, “painterly”).

In-painting & masked edits

• Upload or select an existing Create output, draw a mask, and provide a new prompt to modify only that region.

Advanced parameter panel (and API parity)

```

{

"seed": 1234,

"steps": 30,

"guidance": 7.5,

"aspect": "3:2"

}

```

Implement a TRUE conversational loop

• Use the LLM to reason about the user’s feedback (“make the lighting moodier, keep the same subject”) and translate it into the next diffusion call, preserving prompt history.

Multimodal recursion

• Allow chaining: Create → Vision-LLM analysis → updated prompt → Create, until the user is satisfied.

There will be several benefits to that:

• Faster, finer-grained creative iterations.

• Better alignment with brand/style guidelines.

• Keeps users inside the You.com ecosystem instead of exporting to third-party editors.

• Showcases You.com’s multimodal strengths, differentiating it from single-purpose generators.

June 17, 2025

Julien

Sami (from You.com), could you eventually merge my demand with the rvkproject 's post dated from November the 7 which title was "Customize LLM Parameters (Temperature, Top P, etc.) for Custom Agents"?
Thanks!
Julien

Julien

Friendly note to the team
And yes, I totally get that Create
 might feel like a “nice-to-have” compared to core search & chat. You probably added a straightforward diffusion model to meet market demand and to keep the product sheet shiny—fair enough, that’s smart from a marketing standpoint! 😄  
But for many of us, visual generation isn’t secondary at all. Unlocking deeper controls would turn Create
 from a cool demo into a daily, professional tool—and showcase the full multimodal power that makes You.com unique.