Skip to content

AI providers

Settings → AI Providers (/admin/settings/ai/providers) is where the LLM vendor credentials live: API keys and models for the server-side eval judge, plus API keys for the embedding providers used by judge-fact deduplication. This page is credentials only — it doesn’t pick which vendor or provider is active; that happens on Evals.

Configure one or both vendors to enable the server-side LLM-as-judge runner. When both are configured, users pick the vendor at run time.

KeyDescription
Evals:ServerRunner:Anthropic:ApiKeyThe Anthropic API key. Stored encrypted at rest.
Evals:ServerRunner:Anthropic:ModelThe judge model, e.g. claude-sonnet-4-6 or claude-opus-4-7.
KeyDescription
Evals:ServerRunner:OpenAI:ApiKeyThe OpenAI API key. Stored encrypted at rest.
Evals:ServerRunner:OpenAI:ModelThe judge model, e.g. gpt-4o.

Available on the Team and Enterprise plans. On other plans, bring your own OpenAI or Anthropic key above.

The API token, endpoint, and gateway are managed by Kurrent (set at deployment) — only the model is overridable:

KeyDescription
Evals:ServerRunner:KurrentAI:ModelThe judge model.

Hosted or on-prem:

  • Hosted Cohere API — set just the API key.
  • On-prem / airgapped deployment — set the base URL to your API base, e.g. https://cohere.internal (omit /v2/chat; it’s appended). The key is then optional.
KeyDescription
Evals:ServerRunner:Cohere:BaseUrlBlank uses the hosted SaaS API. Set for an on-prem deployment.
Evals:ServerRunner:Cohere:ApiKeyRequired for the hosted API; optional on-prem.
Evals:ServerRunner:Cohere:ModelThe judge model.

API keys for the embedding providers used by the judge-fact semantic deduplication pipeline. Select which provider is active on the Evals page.

KeyDefaultDescription
Embedding:Voyage:ApiKeyVoyage API key. Required when judge-fact dedup is enabled.
Embedding:Voyage:Modelvoyage-4-liteVoyage embedding model.
KeyDefaultDescription
Embedding:OpenAI:ApiKeyOpenAI API key. Required when judge-fact dedup is enabled.
Embedding:OpenAI:Modeltext-embedding-3-smallOpenAI embedding model.
Embedding:OpenAI:DimensionsDimension count, e.g. 1536.

Estimated cost rates sync automatically from models.dev — no configuration needed. The page shows the pricing source, the count of priced models, and the last sync time, with a warning if the last successful sync is more than 48 hours old.

  • Evals — selects the active embedding provider and sets server-runner run limits.
  • Assistants — reuses these same vendor credentials.