[Feature request] Allow users to select the AI completion model per chat

It would be very useful if Docmost could support multiple configured AI completion models and allow users to select which model should be used for a chat.

At the moment, the completion model is configured globally through:

AI_COMPLETION_MODEL=...

Changing the model therefore requires changing the configuration and restarting/recreating the Docmost container.

For self-hosted setups using Ollama, it is common to have several models installed at the same time. Different models can perform better for different tasks, especially in technical knowledge bases and RAG workflows.

Proposed functionality

An admin should be able to define which completion models are available in Docmost.

For example:

qwen3:8b
qwen2.5:7b
llama3.1:8b
gemma3:4b

The user could then select the desired model directly in the AI Chat, similar to model selection in ChatGPT, Claude or other AI chat applications.

Ideally, the configuration could support:

• A default completion model for the workspace
• An admin-defined allowlist of available models
• A model selector in AI Chat
• The selected model stored per chat
• The option for admins to hide the selector from normal users
• A fallback model if the selected model becomes unavailable

Why this would be useful

Different completion models can behave very differently with the same retrieved context.

For example, one model may:

• answer technical questions more precisely
• follow RAG context better
• perform better in German or other languages
• be faster for simple questions
• handle larger contexts better
• be better suited for complex reasoning

Being able to switch models directly in the chat would also make it much easier to compare model quality without restarting Docmost.

Self-hosted / Ollama use case

This would be especially useful for local Ollama installations, because multiple models can already be installed on the same Ollama instance.

Docmost could potentially retrieve the available models from Ollama and let the administrator decide which ones should be exposed to users.

Example

Instead of configuring only:

AI_COMPLETION_MODEL=qwen3:8b

the administrator could configure several allowed models and define:

Default: qwen3:8b

Available:
- qwen3:8b
- qwen2.5:7b
- llama3.1:8b
- gemma3:4b

The user could then choose the model from a dropdown in AI Chat.

This would make Docmost much more flexible for self-hosted AI setups and would be particularly useful for testing and optimising RAG performance.