# Settings: creativity, diversity threshold, reasoning effort and verbosity

[*AI model*](https://manual.selarti.com/manual-en/additional/ai-models) *tuning can vary across a number of parameters. This article explains them in more detail.*

Parameters **creativity** and **diversity threshold** determine the degree of randomness and creativity in the AI agent's responses. They are interconnected but have different approaches and principles of operation.

#### 📌 **Creativity**

**What is it**: A parameter that controls the randomness of text generation.

**How it works**:

* Values from `0` to `2` (usually used from `0` to `1`).
* The higher the value, the more random, unexpected, and creative the AI's responses are.
* The lower the value, the more conservative, consistent, and predictable the responses.

**Examples**:

* `0` — almost identical response to similar queries (clarity and predictability).
* `0.7` – `1.0` — more diverse and creative responses.
* `1.5` — maximum creativity and unexpected associations (but also more mistakes).

#### 📌 **Diversity threshold**:

**What is it**: A parameter that limits the pool of the most probable tokens considered by the model when generating text.

**How it works**:

* Values from `0` to `1`.
* If you set the value, for example, `0.1`, then the model will consider only the most probable tokens whose total probability is 90%.
* This means that less probable options are completely ignored, and more probable tokens remain, but within this threshold.

**Examples**:

* `diversity threshold = 0.1` — only the most probable words (limited and predictable responses).
* `diversity threshold = 0.9 – 1` — allows generating more diverse and creative text.

***

#### 🚩 **What is the difference and how to use**:

* **Creativity** directly regulates the randomness of the response (like a «chaos slider»).
* **Diversity threshold** limits the range of response options by probabilities (like a «best options filter»).

Usually, for assistant settings, **one** of these parameters is used, not both at the same time, because they act on the same token selection process.

**🎯 Usage recommendations:**

* If you need accuracy and stability — `creativity` close to `0` (for example, 0.1–0.3).
* If originality, liveliness, and naturalness are important — `creativity` from `0.7` to `1`.
* If you need a balance between randomness and stability with a guarantee of response quality — `diversity threshold` ≈ `0.9`.

It is usually recommended to choose one of the parameters, most often **creativity**, and **diversity threshold** — is an alternative approach if you want to control diversity through probabilities, not just the level of randomness.

***

#### 📌 Reasoning effort

**What is it**: An AI operating parameter that simulates the processes of human thought and reasoning.

**Can take values**:

* «Low» — thinks little, answers quickly and cheaply.
* «Medium» — the golden mean (standard).
* «Hight» — thinks for a long time, expensively, but in the end gives the most intelligent answer.

#### 📌 Verbosity

**What is it**: An AI operating parameter that determines how detailed a response a user will receive in a conversation with the AI.

**Can take values**:

* «Low» — answers briefly, quickly and cheaply.
* «Medium» — the golden mean (standard).
* «Hight» — thinks longer and gives a detailed answer.

***

**🎯 Usage recommendations:**

By default, we recommend using the «Medium». If response speed is your top priority, choose «Low». If you expect the AI ​​to perform auditing, training, and transfer complex information, choose «Hight».