How it works
Two separate, independent roles: generating answers (chat) and vectorisation (embeddings). You can assign them to different providers (mixed mode).
The options
- Claude (via the DGT proxy) — the recommended default for generation.
- Mistral (EU sovereignty), OpenAI, Google Gemini — with the client's own key, configurable per space.
- Local / open source (Ollama…) — self-hosted, no data leaves the server.
Configure
Open
Administration → AI provider configuration.Choose the provider
Globally or per client space (key stored encrypted).
Local vectorisation (optional)
For a sovereign RAG, enforce local embeddings (Ollama server), while keeping generation on Claude.
- Anthropic has no embeddings API: vectorisation goes through Mistral / OpenAI / Gemini or a local model, even when generation stays on Claude.
- The companion uses a low-cost model by default (Haiku).