Sourced answers (RAG)
Hybrid search, contextual enrichment and verified citations: every answer points to the excerpts that support it.
Language models
Choose your models, stay in control
OpenAI, Mistral, Anthropic, Gemini or a model you host yourself: a choice per assistant, encrypted keys and a benchmark to decide based on facts.
Supported providers
OpenAI Mistral Anthropic Gemini Ollama vLLM OpenAI-compatible
Same questions, multiple models
For each answer: text, tokens used and latency; Excel export.
The problem
Why it matters
The model landscape changes every month. Tying your assistants to a single provider exposes you to cost increases, behavior changes or confidentiality constraints.
You need to be able to switch models per assistant, compare them objectively and protect access keys.
In practice
- Register credentialsA provider key, or the address of a self-hosted model server, is stored encrypted.
- Choose per assistantProvider, model, temperature, response length, reasoning, JSON output.
- CompareThe benchmark runs the same questions on several models or versions and measures tokens and latency.
Capabilities
What you can do
Each capability is shown with its actual availability.
- OpenAI, Mistral, Anthropic, Google GeminiAvailable
- Ollama, vLLM and OpenAI-compatible servicesSo answers can be generated by a model you host yourself.Available
- Live model list and connection testAvailable
- Extended reasoningThinking budget (Anthropic), reasoning effort (OpenAI).Available
- Structured JSON outputWith an optional JSON schema.Available
- Encrypted credential vault (AES-256-GCM)Sharing with named users, or availability with a daily cap.Available
- Multi-model and multi-version benchmarkTokens and latency per question, Excel export.Available
- Fully offline operationEmbedding, transcription and OCR currently go through OpenAI or Mistral.Experimental
Benefits
What you gain
No vendor lock-in
Switch models without rebuilding your assistants or your knowledge bases.
Evidence-based choices
Compare quality, latency and usage on your own questions.
Protected keys
Encrypted at rest, never exposed to end users.
Requirements and limitations
Related features
AI assistants
One assistant per need: role, tone, knowledge, tools and publishing rules, configured from a single interface and versioned.
Security and governance
Roles, SSO including FranceConnect+, white-label organizations, GDPR retention, audit log, content validation and migration between environments.
Next step
Cloud, in-house model servers: what each option means in practice.