> ## Documentation Index
> Fetch the complete documentation index at: https://docs.verba.ink/llms.txt
> Use this file to discover all available pages before exploring further.

# AI engine settings

> Tune model behavior, response length, context, and search.

## Questions this guide answers

* Which settings affect creativity vs consistency?
* Why are responses short or long?
* How much context and max tokens can my plan use?
* What does web search actually do?

## Core controls

| Setting       | Range          | What it changes                    |
| ------------- | -------------- | ---------------------------------- |
| Temperature   | `0..2`         | Randomness and creativity          |
| Top-p         | `0..1`         | Diversity of token selection       |
| Model Context | plan-limited   | Number of recent messages included |
| Reply Style   | preset         | Voice/format tendency              |
| Multi-message | on/off + delay | Sends split follow-up replies      |
| Web search    | on/off         | Allows live web-grounded replies   |

## System instructions (behavior prompt)

System instructions are your verb's persistent behavior rules.

Where to set:

* Dashboard -> Bot -> AI Engine -> Behavior
* Field name: `systemInstructions`

Limit:

* Up to `8000` characters

How they interact with other layers:

* `systemInstructions`: behavior/rules/tone constraints
* Training examples: style shaping and response pattern hints
* Long-term memory: durable facts/preferences
* Knowledge entries: factual/reference content
* Conversation context: recent turns in current chat/thread/session

Practical rule:

* Put "how to behave" in system instructions.
* Put "facts to remember" in knowledge/memory.
* Put "how to phrase outputs" in training examples.

## Plan-based limits

Current default limits:

| Plan  | Max model context | Max response tokens |
| ----- | ----------------- | ------------------- |
| Free  | `50`              | `4096`              |
| Plus  | `75`              | `8192`              |
| Pro   | `100`             | `16384`             |
| Ultra | `150`             | `32768`             |

<Note>
  If you request settings above your tier limits, they are clamped or rejected
  depending on the endpoint.
</Note>

## Premium AI model access

AI model availability is plan-based.

If you open the AI Engine and choose a model that is outside your current tier,
Verba shows an upgrade prompt with the number of additional premium AI models
available above your current plan.

That upgrade hint is dynamic, so the exact count can differ depending on:

* Your current plan
* The active model catalog
* Whether new premium models have been added since your last visit

## Recommended presets

<CardGroup cols={2}>
  <Card title="Reliable support" icon="shield">
    Temperature `0.4-0.7`, top-p `0.7-0.9`, moderate context.
  </Card>

  <Card title="Creative roleplay" icon="sparkles">
    Temperature `0.8-1.1`, top-p `0.9-1.0`, higher context.
  </Card>

  <Card title="Cost-aware" icon="wallet">
    Lower context, web search off, shorter reply style.
  </Card>

  <Card title="Fast troubleshooting" icon="bolt">
    Lower context + deterministic settings to reduce latency variance.
  </Card>
</CardGroup>

## Multi-message behavior

When enabled:

* Verba can split a response into multiple shorter messages.
* Delay between parts is configurable (`0..10000ms`).

Use this for natural chat pacing; disable if you want one compact answer.

## Web search behavior

When enabled, Verba may perform model-driven search planning before the final answer.
This improves freshness, but can increase:

* Latency
* Token usage
* Cost

Use web search for:

* News
* Live pricing
* Fast-changing product details

Keep it off for:

* Roleplay
* Stable canon/lore
* Deterministic support flows

## System instructions best practices

* Keep instructions concrete and scoped.
* Avoid contradictory rules.
* Prefer short imperative bullets over long prose.
* Include failure policy (what to do when unknown).
* Specify output format explicitly (headings, bullets, steps, code blocks).

## Quick diagnostics

<AccordionGroup>
  <Accordion title="Replies are too generic">
    Lower temperature, tighten system instructions, and add targeted training examples.
  </Accordion>

  <Accordion title="Bot forgets recent details">
    Increase model context (within your plan limit).
  </Accordion>

  <Accordion title="Replies are too short">
    Increase max tokens and use a fuller reply style.
  </Accordion>

  <Accordion title="Replies are too slow">
    Reduce context and disable web search unless needed.
  </Accordion>
</AccordionGroup>

<Card title="Bot Settings Reference" icon="book" href="/guides/bot-settings-reference">
  See all AI-related fields and limits in one place.
</Card>