--- title: "Understanding AI Model Settings" slug: "understanding-ai-model-settings" tags: ["unqorkai"] status: "update" updated: 2026-10-26T21:40:33Z published: 2026-10-26T21:40:33Z canonical: "docs.unqork.io/understanding-ai-model-settings" --- > ## Documentation Index > Fetch the complete documentation index at: https://docs.unqork.io/llms.txt > Use this file to discover all available pages before exploring further. # Understanding AI Model Settings AI components expose settings that control how the AI model generates its responses. Which models are available, and which settings a model reads, depend on the LLM credential the component runs on. > **[Chat Agent Smart Component](/unqorkai/docs/chat-agent-smart-component)**: These settings also apply to the Chat Agent Smart Component, but they are configured in **Settings > LLM Configs** inside the LLM Config rather than directly on the component itself. ## Instructions Instructions tell the AI model its role, tone, and constraints before the conversation or task begins. > [!NOTE] > In AI terminology, this field is also called a system prompt or system instruction. The component label is **Instructions**. **Example instructions:** > You are a customer support specialist for a retail brand. Answer questions politely and concisely. If you do not know the answer, say so and suggest the customer contact support. > You are a document summarizer. Summarize the provided text in three bullet points. Focus on key decisions, action items, and deadlines. **Tips for writing effective instructions:** - Be specific about the role and tone. - Define what the model should do when it does not know an answer. - If output format matters, specify it: bullet points, paragraph, numbered list, and so on. ## Temperature Temperature sets the variability of the model's output. Lower values produce more focused, consistent responses; higher values produce more varied, creative ones. The UI exposes a 0–2 scale. Most models do not honor the full range; behavior varies by model. | Value | Effect | | --- | --- | | **0–0.3** | Responses are more focused, consistent, and predictable. Use for factual Q&A, document retrieval, or any context where accuracy matters more than variety. | | **0.4–0.8** | Balanced responses. Good for most use cases. | | **0.9–2** | Responses are more varied and creative. Use for brainstorming or scenarios where novelty is valuable. | The AI Summarizer defaults to `0.5`. **Model behavior varies.** The 0–2 range applies as configured only for Gemini 3.x and Nova 2 Lite. Other models handle temperature differently: | Model | Temperature behavior | | --- | --- | | Gemini 3.x, Nova 2 Lite | Applied as configured (0–2) | | GPT-5.x, Haiku 4.5 | Pinned to 1.0 — any value is silently clamped | | Opus 4.8, Sonnet 5, Fable 5 | Temperature not supported — the parameter is ignored | The UI accepts 0–2 regardless of model and clamps silently where needed. ## Thinking Level Thinking Level is a simplified control for reasoning depth. Which setting the model reads depends on the model; the Creator does not choose between them. | Level | Effect | | --- | --- | | **Off** | No internal reasoning. Fastest responses. | | **Medium** | Moderate reasoning. Good for longer documents or content that requires nuance. | | **High** | Extended reasoning. Use for complex content where accuracy is critical. | **Which models read Thinking Level:** | Setting read | Models | | --- | --- | | **Thinking Level** | GPT-5.x, Gemini 3.x, Opus 4.8, Sonnet 5 | | **Thinking Token Budget** | Haiku 4.5 only | | **Neither (both ignored)** | Fable 5, Nova 2 Lite | ## Thinking Token Budget The Thinking Token Budget controls how much internal reasoning the model performs before generating a response. This setting applies only to Haiku 4.5. On all other models it is ignored. Values below 1024 are clamped to 1024, giving a practical range of 1024 to 4096. | Value | Effect | | --- | --- | | **~1024** | Minimal internal reasoning. Fast responses, suitable for straightforward tasks. | | **~2048** | Moderate reasoning. Good for questions that require the model to consider multiple factors. | | **~4096** | Extended reasoning. Use for complex multi-step problems or scenarios where accuracy is critical. | > [!NOTE] > The AI Summarizer defaults to `1024`. Increasing the Thinking Token Budget increases processing time and cost. > **[Chat Agent Smart Component](/unqorkai/docs/chat-agent-smart-component)**: For the Chat Agent, the LLM Config handles thinking settings and replaces the need to set the Thinking Token Budget directly in most use cases. ## Max Output Tokens Max Tokens sets the maximum length of the model's response. One token is roughly equivalent to four characters of text. | Value | Approximate output | | --- | --- | | **1024** | Several paragraphs | | **2048** | A detailed multi-paragraph response | | **4096** | A long, comprehensive response | The AI Summarizer defaults to `2048`. The maximum value varies by model: | Model | Max output tokens | | --- | --- | | Gemini models | 65,535 | | Sonnet 5, Haiku 4.5, Nova 2 Lite | 64,000 | | GPT-5.x, Opus 4.8, Fable 5 | 128,000 | Values above the model's maximum are clamped to that maximum. If responses are being cut off mid-sentence, increase this value. If responses are unnecessarily long, decrease it. > [!NOTE] > Max Output Tokens sets a ceiling, not a target. The model generates as much text as needed up to the limit. ## Choosing Settings Most use cases work well with the default settings. Adjust when you notice the following issues: | Issue | Adjustment | | --- | --- | | Responses feel generic or repetitive | Increase Temperature (Gemini 3.x and Nova 2 Lite only) | | Responses are inconsistent or off-topic | Decrease Temperature (Gemini 3.x and Nova 2 Lite only) | | The model struggles with complex content | Increase Thinking Level (GPT-5.x, Gemini 3.x, Opus 4.8, Sonnet 5) or Thinking Token Budget (Haiku 4.5) | | Summaries lack depth or miss nuance | Increase Thinking Level | | Responses are cut off before finishing | Increase Max Output Tokens | | Responses are unnecessarily long | Decrease Max Output Tokens | --- **Changelog** | Date | Change | | --- | --- | | 2026-09-12 | Rewrote with model-specific behavior for Temperature, Thinking Level, Thinking Token Budget, and Max Output Tokens; added model-compatibility table; renamed System Prompt to Instructions; added AI Agent LLM Config notes; fixed incorrect "retired AI Agent" reference. | | 2026-08-19 | Clarified that available models and applicable settings come from the LLM credential. | | — | Initial publication. | Large Language Models (LLM) are artificial intelligence (AI) programs that use deep learning algorithms to analyze and generate text and other content based on large amounts of data. A key-value pair is a fundamental data structure where a unique "key" is associated with a specific "value," used for organizing and retrieving data efficiently, like in dictionaries or databases. The UI (user interface) is a device's end-user interaction and communication point. It can include display screens, keyboards, and mouse. It is also how a user interacts with your application. Also known as Unqork Users, or Designer Users; is anyone who is inside the Unqork platform.