Adaline Evaluations

Concepts

Evaluations run a prompt against a dataset and score each row with one evaluator. They are asynchronous: create a run, poll its status, then read paginated results.

Key terms:

• Evaluation — one run, identified by runId
• Evaluator — the scoring configuration, identified by evaluatorId
• Dataset — rows that provide prompt inputs and optional expected values
• Grade — pass, fail, or unknown
• Metrics — aggregate pass/fail/unknown counts, cost, latency, and token count

Status Lifecycle

queued -> running -> completed
                  -> failed
                  -> cancelling -> cancelled

Configuration

Set these environment variables when credentials are available:

• ADALINE_API_KEY — workspace API key from Admin > API Keys
• ADALINE_PROMPT_ID — prompt to evaluate
• ADALINE_EVALUATOR_ID — evaluator to run
• ADALINE_DATASET_ID — optional dataset override

Base URL: https://api.adaline.ai/v2

Quick Triage

Running an Evaluation

Step 1 — Create run

curl -X POST "https://api.adaline.ai/v2/prompts/$ADALINE_PROMPT_ID/evaluations" \
  -H "Authorization: Bearer $ADALINE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "evaluatorId": "evaluator_abc123",
    "datasetId": "dataset_abc123"
  }'

The response returns runId. Use that value as evaluationId in status/results/cancel calls.

Step 2 — Poll status

curl "https://api.adaline.ai/v2/prompts/$ADALINE_PROMPT_ID/evaluations/$RUN_ID" \
  -H "Authorization: Bearer $ADALINE_API_KEY"

Step 3 — Fetch results

curl "https://api.adaline.ai/v2/prompts/$ADALINE_PROMPT_ID/evaluations/$RUN_ID/results?grade=fail&expand=row&limit=50" \
  -H "Authorization: Bearer $ADALINE_API_KEY"

Step 4 — Cancel if needed

curl -X POST "https://api.adaline.ai/v2/prompts/$ADALINE_PROMPT_ID/evaluations/$RUN_ID/cancel" \
  -H "Authorization: Bearer $ADALINE_API_KEY"

SDK Usage

const run = await adaline.prompts.evaluations.create({
  promptId,
  evaluation: { evaluatorId, datasetId },
});

const status = await adaline.prompts.evaluations.get({
  promptId,
  evaluationId: run.runId,
});

const results = await adaline.prompts.evaluations.results.list({
  promptId,
  evaluationId: run.runId,
  grade: 'fail',
  expand: 'row',
});

run = await adaline.prompts.evaluations.create(
    prompt_id=prompt_id,
    evaluation=CreateEvaluationRequest(evaluator_id=evaluator_id, dataset_id=dataset_id),
)

status = await adaline.prompts.evaluations.get(
    prompt_id=prompt_id,
    evaluation_id=run.run_id,
)

results = await adaline.prompts.evaluations.results.list(
    prompt_id=prompt_id,
    evaluation_id=run.run_id,
    grade="fail",
    expand="row",
)

Best Practices

1. Use one evaluator per run; create multiple runs when you need multiple evaluators.
2. Persist runId in CI or job metadata so later steps can poll and fetch results.
3. Poll status with backoff; do not tight-loop.
4. Gate deploy/promotion on terminal status and acceptable metrics.
5. Inspect failing rows with grade=fail&expand=row.

References

See references/api.md for request/response schemas and curl examples.