Evals and scores

Publish Anvia evaluation metrics with createLangfuseEvalReporter() or combine an evaluation and Langfuse dataset run with runEvalAsExperiment().

const reporter = createLangfuseEvalReporter(tracing, {
  onMissingTrace: 'warn',
  includeMessages: false,
})

Numeric, categorical, and boolean outcomes preserve their Langfuse data type. Metric configuration, metadata, trace references, and input or expected summaries are included according to reporter options.

Direct scores

await tracing.score({
  traceId,
  observationId,
  name: 'quality',
  value: 1,
  dataType: 'NUMERIC',
})

Direct sending is the default. Set scoreBatchSize to enable the in-memory queue; it drains by size, interval, flushScores(), flush(), or shutdown(). Retryable 429 and 5xx responses use bounded exponential backoff.

If retries are exhausted, LangfuseScoreError.scores contains the failed payloads. The queue is process memory, not durable storage, so use a job system when score delivery must survive restarts.