Skip to main content

Language Models (LLM)

From Manager → Settings → AI Models, configure the language models usable by agents, the supervisor, the switch, interceptors and benchmarking.

A model always relies on a Provider — that is, an access (API key, endpoint) configured beforehand by an administrator. If no suitable provider appears in the list, ask an administrator to add one.

Configuring a model​

  • Name — label shown in model selectors.
  • Provider — chosen among the LLM-type providers configured by the administrator.
  • Model name — technical identifier, e.g. gpt-4o, claude-sonnet-5.
  • Token limit — default 4096.
  • Inference parameters — temperature, top-p, top-k, reasoning (thinking), each individually toggleable.
  • Advanced initialization parameters — additional technical options passed to the model's client (JSON).
  • Active — an inactive model no longer appears in selectors.
  • Default general assistant — only one model can carry this flag at a time (enabling it on one model automatically removes it from the others); this is the model used, for example, as the default judge for benchmarking.

A "Test" button actually calls the model with a minimal prompt to validate the connection, and a "Probe" button checks its capabilities (token limit, reasoning support…) and pre-fills the suggested settings. Deleting a model still used by agents or pipelines requires designating a replacement model.

Cost per token​

Pricing (price per million tokens, input and output) is not set on this screen, but from the Manager → AI Cost Tracking tab.