AI guide
These settings shape how Iris chats when someone tags her, replies to her, or uses ask. A manager changes them on the dashboard, next to the provider, the model, and the key.
Temperature
Temperature controls how varied the wording is. Low values give steady, predictable answers that are good for facts, code, and math. High values give more varied and playful answers, and at the top of the range they start to wander. Leave it empty to use the model's own default, which is the right choice for most servers.
Many newer models refuse a custom temperature. When a model refuses it, Iris asks again without it and remembers that for the model, so the question is still answered with the default. While reasoning is on, Iris does not send a temperature at all, because reasoning models do not take one.
| Provider | Range | Default | Advice |
|---|---|---|---|
| OpenAI | 0 to 2 | 1.0 | 0.7 to 1.0 for chat. GPT-6 models take it only while reasoning is off. GPT-6.1 Sol and GPT-6 Astra never take it. |
| Anthropic | 0 to 1 | 1.0 | Leave it on the default. Claude Haiku 5.5, Sonnet 5, Opus 4.7, and newer refuse any custom value. Older models such as Claude Haiku 4.5 take 0 to 1 while reasoning is off. |
| Google Gemini | 0 to 2 | 1.0 | Keep 1.0 for Gemini 3. Google warns that values below 1.0 can make Gemini 3 loop or answer worse. Gemini 2.5 models take 0 to 2. |
| OpenRouter | 0 to 2 | 1.0 | Follow the advice for the model's own provider. OpenRouter passes the value on. A model that refuses it, such as a recent Claude, answers with its default. |
| Groq | 0 to 2 | 1.0 | 0.6 to 1.0 for chat. Most open models on Groq take it. |
| Mistral | 0 to 1 | set per model | Mistral recommends 0 to 0.7. Above 0.7, answers get random quickly. |
| DeepSeek | 0 to 2 | 1.0 | DeepSeek recommends 1.3 for conversation, 1.5 for creative writing, and 0 for code and math. Reasoning models ignore it. |
| xAI | 0 to 2 | 1.0 | 0.7 to 1.0 for chat. Some Grok reasoning models may refuse it. |
Reasoning
With reasoning on, the model thinks before it answers. That helps with hard questions, but answers come slower and cost more, because the thinking is billed as output. It is off by default. Some models always reason, and Iris then uses the lightest setting that model accepts.
Personality
An optional line of at most 300 characters, such as "playful and a little sarcastic". It sets who Iris is and how she talks, so she answers questions about herself from it. Iris still answers in the language of the question, keeps answers short, and follows her safety rules, which a personality cannot change.
Web research
With a search provider and its key saved next to the AI key, research and the chat tool can answer with current facts instead of model memory. A manager picks Tavily, Brave Search, or Serper. Iris checks that key with one small test search, then encrypts it. The question goes to that provider, the top results come back, and the AI provider writes one short answer from them in this server's personality. When the question asks for sources or links, up to three real result links follow that answer; she never invents a URL. Without the search key, research stays off. The lookups that already exist, such as news, hn, and wiki, do not use it.
What is kept
By default Iris stores one row per answer: the question, the reply Discord received, the channel, and a few review flags, with secrets redacted and both texts capped. It exists so the chat guards can be reviewed, it is never shown to other members, and rows older than 90 days are deleted. A manager can turn the switch off on the dashboard, which stops new rows. Nothing else in the chat is stored.
OpenRouter
Add :nitro to a model name, such as anthropic/claude-haiku-5.5:nitro, to use the fastest host for that model. OpenRouter's privacy settings decide which hosts may store or train on prompts.