AI models
Connect AI providers, register provider models and publish deployments for Chouse AI — sixteen provider types, keys encrypted.
- In the app
- Admin › AI models
/admin/ai-models - Needs
- ai_models:view

Every Chouse AI feature — AI Assist in the editor, the Doctor, in-tab diagnosis, incident explanations, watchers, the operational brief — runs on a model you configure here. Nothing is sent to a provider until you add one.
Setup has three layers, one tab each:
| Tab | What it holds |
|---|---|
| Providers | Who you call and with which credentials |
| Provider models | Which model of that provider (its exact model ID) and its runtime parameters |
| Deployments | The named configurations people pick in the app, each pointing at a provider model; one is the default |
1. Add a provider
Providers › Add provider (needs ai_models:create): a display name, a provider type, and its credentials.
| Provider type | Needs |
|---|---|
| OpenAI, Anthropic, Google, Groq, Mistral, Cohere, xAI, DeepSeek, Cerebras, Fireworks, Together, OpenRouter | API key |
| OpenAI-compatible | Base URL and API key (vLLM, LiteLLM, LM Studio, …) |
| Azure OpenAI | Resource endpoint (https://<resource>.openai.azure.com) and API key; model IDs are your Azure deployment names |
| Amazon Bedrock | AWS region, access key ID and secret access key |
| Ollama | Base URL (no key) |
Credentials are stored encrypted and never returned to the browser; to rotate them, enter them again. A provider can be switched Inactive without deleting it.
2. Add a provider model
Provider models › Add base model: pick the provider and enter the provider model ID exactly as the provider expects it (e.g. gpt-4o). Optional runtime parameters — stop sequences, API version, Gemini safety settings, extra request parameters as JSON — apply to every deployment that uses the model; empty fields keep the defaults.
Prices
Each provider model can have a price in USD per 1M tokens: Input, Output, and optionally Cached input and Cache write (they default to the input price). AI activity uses it to estimate cost. Leave every price empty to show tokens only; enter 0 for a free, self-hosted model. A new price applies to requests from then on. The Price per 1M column shows input / output.
Chat attachments
Chat attachments says what users may attach in the chat besides text files: Images and PDFs, for models that read them. Leave both off for a text-only model. Images and PDFs aren't scanned by guardrails. See Using the chat.
3. Create a deployment
Deployments › Add configuration: a friendly deployment name people will see, and the provider model it uses. Mark it Active, and Set as default for the deployment used when nobody picks one.
Models and agents
Every AI feature runs on an AI agent. An agent can name one of these deployments as its model; one that doesn't uses the default deployment, and a model the user picks in the chat or SQL editor wins over both. On Data, Dashboards (Ask AI) and AI Governance, the model button in the header is each user's own pick for every AI feature there; System default uses the default deployment. If the picked deployment is later removed or deactivated, the button says Unavailable and AI features fail until you pick another, rather than quietly using a different model.
Context window
Agents summarize a long conversation before it overflows the model: at 85% of the model's input window they replace older messages with a summary and keep the latest 10%. CHouse knows the window of models in LangChain's built-in model list (the form shows it as from model profile). For any other model — most OpenAI-compatible and local models — set Context window (tokens) on the provider model; you can't save or run such a model without it. Set it for a known model too, to override the built-in value. Max output tokens must be smaller than the window.
The Reliability & agent runtime parameters of a provider model — Recursion limit and Run timeout (ms) — override the step and time limits of every agent run on that model. Leave them empty to use each agent's own tuning.
Keep prompts in your network
To keep prompts and schema context inside your network, use Ollama or any OpenAI-compatible server you host.
Who can use AI
Configuring models needs the ai_models:* permissions. Using AI is separate: ai:optimize for optimize/debug/diagnose and the Doctor's explanations, ai:chat for chat, and doctor:run to start Doctor scans. Changes to providers, models and deployments, and every Doctor scan, are recorded in the audit log.