Runtime scoring

LensClient.score() records a score against an existing Lens trace. It uses the client's owned OTLP logs provider, so applications can make production feedback observable without configuring a separate evaluation suite or OpenTelemetry logger.

await lens.score({
  id: feedbackId,
  traceId,
  observationId,
  responseId,
  name: 'user-feedback',
  value: liked ? 1 : 0,
  dataType: 'BOOLEAN',
  source: 'end_user',
  comment,
  metadata: { channel: 'thumbs' },
})

Lens v0.9.1 or later preserves end_user provenance and displays it separately from telemetry-generated and internal human evaluations. The API is not limited to like/dislike:

dataType	Accepted value	Example
BOOLEAN	1 or 0	Helpful/not helpful
NUMERIC	A finite number	A 1–5 rating or confidence score
CATEGORICAL	A string	A feedback reason or rubric category

For Boolean scores, 1 infers a pass outcome and 0 infers fail. Supply outcome and label when the product uses different semantics.

Make feedback durable and attributable

• A valid 32-character traceId is required. Add the optional 16-character observationId to target a particular span, and responseId when feedback applies to a model response.
• Use a stable opaque id when a later vote should replace or supersede the same application feedback record. Omit it when every submission must remain a separate event.
• Set source: 'end_user' only after authenticating and authorizing the feedback request. Lens provenance is an observability classification, not proof of user identity or permission.
• Keep tenant and user ownership in application data. If metadata needs an identifier, prefer an approved opaque or hashed value over an email address or other raw personal data.

score() resolves after the client has queued the log record; it does not force exporter delivery. Call lens.flush() at a short-lived delivery boundary and lens.close() during graceful shutdown. A disabled optional client treats scoring as a no-op, while a closed client rejects new calls.

The client rejects invalid trace or observation IDs, empty names and explicit IDs, non-finite numbers, values that do not match dataType, and comments longer than 2,000 characters.