Skip to content

Evals

Settings → Evals (/admin/settings/ai/evals) configures how the server-side LLM-as-judge runner behaves once judge-vendor credentials are in place: which judge vendor and model to use, how long a run is allowed to take, how much trace it sees, which repos auto-evaluate, and whether judge findings get deduplicated with embeddings. Judge-vendor API keys live on AI Providers — this page picks the active judge vendor and sets behavior, not credentials.

Selects the judge vendor and model the server-side runner uses by default. Credentials for the chosen judge vendor come from AI Providers.

KeyTypeDefaultDescription
Evals:Agent:Providerenum Anthropic / OpenAI / KurrentAI / Cohere / OpenRouterAnthropic (runtime)Default judge vendor for server eval runs.
Evals:Agent:Modelstringclaude-sonnet-4-6Default judge model.

Both fields render blank in the form and fall back at runtime:

  • Provider — an unset value falls back to the legacy Evals:AutoEval:Provider if present, otherwise Anthropic.
  • Model — an unset value falls back to the selected judge vendor’s configured judge model, otherwise that judge vendor’s built-in default.

Per-run budgets for the server-side runner:

KeyRangeDefaultDescription
Evals:ServerRunner:RunTimeoutSeconds60–36001800 (30 min)Wall-clock budget for an entire eval run. On expiry the run surfaces as EvalFailed("cancelled").
Evals:ServerRunner:JudgeTimeoutSeconds30–3600300 (5 min)Wall-clock budget for a single judge call.
Evals:ServerRunner:RetrospectiveReserveSeconds0–900120Seconds reserved out of the run timeout for the retrospective synthesis pass.
Evals:ServerRunner:MaxTraceChars10000–2000000400000 (~100k tokens)Upper bound on the formatted trace passed to each judge prompt. When exceeded, the orchestrator keeps the most recent events and prepends a marker noting how many were elided.

A table of the repositories visible to you, each with two per-repository overrides:

  • Auto-evaluate on close — runs an eval automatically when a session on this repo closes. Disabled until a server eval runner is configured (see the Evals agent above).
  • Fact-promotion threshold (integer ≥ 1) — the minimum weight a judge fact needs before it’s promoted into repo guidelines. The placeholder shows the default minimum weight.

JudgeFactDedup:Enabled is the master switch for the semantic dedup pipeline:

  • Off — eval-generated judge facts are retained verbatim, and no embeddings are computed.
  • On — candidate facts are embedded and compared against the existing pool; near-duplicates collapse into clusters. Toggling this on also triggers a background backfill of any facts that accumulated while dedup was off.

Deduplication requires an active embedding provider with its API key set — configure that on Embeddings.

Evals also has a link out to the eval catalog — the set of evals the server-side runner can execute:

Edit the eval catalog/admin/evals.

  • AI providers — judge-vendor credentials for the runner.
  • Embeddings — the embedding provider that powers judge-fact deduplication.
  • Eval catalog — the question and template library evals draw from.
  • Evaluations — the user-facing feature these settings support.