Telemetry
AI quality

rag_retrieval_evaluated event schema

A versioned RAG evaluation event that separates retrieval scores, answer review, latency, and cost without storing source text, prompts, or generated answers.

Reviewed by the Telemetry product team on . Event grain, ownership, field types, privacy classification, example payload, and validation. Review standards and ownership

Question this contract answers

Which retrieval and prompt versions produce relevant context and grounded accepted answers?

Fields
11
Required
10
Contract evidence path

rag_retrieval_evaluated: from owned boundary to trusted query

Treat rag_retrieval_evaluated as a durable analytical contract: the owner emits one documented grain, a fixture proves the fields, and a reviewed query answers the stated question.

  1. 1

    Outcome becomes final

    RAG evaluation pipeline emits only after the retrieval and answer evaluators finish or the candidate becomes ineligible.

  2. 2

    Contract is bounded

    10 required fields preserve the declared grain: One evaluator result per eligible query, retrieval version, evaluator, and candidate run.

  3. 3

    Fixture is verified

    Check types, UTC time, alternate outcomes, idempotency, and every pseudonymous or review-classified field.

  4. 4

    Question is answered

    Which retrieval and prompt versions produce relevant context and grounded accepted answers?

Grain

One evaluator result per eligible query, retrieval version, evaluator, and candidate run.

Owner

RAG evaluation pipeline

Emit when

After the retrieval and answer evaluators finish or the candidate becomes ineligible.

Field contract

Typed fields with explicit privacy boundaries

Keep existing field names and types stable after production queries depend on them. Optional context should remain bounded, documented, and justified by a specific decision.

FieldTypeRequiredPrivacyMeaning
timestamp_utctimestampyesnon-sensitiveUTC time at the outcome boundary.
event_idstringyesnon-sensitiveStable unique identifier used for deduplication.
releasestringyesnon-sensitiveApplication or service version that emitted the event.
account_idstringyespseudonymousStable internal account identifier, never an email or name.
evaluation_idstringyespseudonymousStable identifier for the evaluation case.
retrieval_versionstringyesnon-sensitiveVersion of the index, retriever, and ranking contract.
evaluator_versionstringyesnon-sensitiveVersioned evaluator and rubric identifier.
retrieval_scorenumberyesnon-sensitiveNormalized retrieval relevance score.
groundedbooleanyesnon-sensitiveWhether the reviewed answer met the grounding rule.
acceptedbooleannonon-sensitiveReviewed downstream acceptance when available.
duration_msnumberyesnon-sensitiveEnd-to-end evaluated workflow duration.

Synthetic JSON event

{
  "timestamp_utc": "2026-07-29T10:12:43Z",
  "event_id": "evt_rag_eval_01",
  "account_id": "acct_8f31",
  "release": "2026.07.3",
  "evaluation_id": "eval_72ab",
  "retrieval_version": "hybrid_v4",
  "evaluator_version": "grounding_rubric_v2",
  "retrieval_score": 0.86,
  "grounded": true,
  "accepted": true,
  "duration_ms": 1280
}

Privacy review

Review identifiers before ingestion

This example uses synthetic identifiers. Pseudonymous values can still be personal data, and review fields can expose business or provider context. Apply your own consent, retention, access, residency, and deletion requirements.

  • account_id: pseudonymous
  • evaluation_id: pseudonymous

Validation checklist

Prove the contract before building a dashboard

  • Send one known rag_retrieval_evaluated fixture after the documented outcome boundary.
  • Verify all 10 required fields arrive with the documented types.
  • Retry the same event identifier and confirm the chosen deduplication behavior.
  • Send a controlled failure or alternate outcome when the workflow supports one.
  • Run the related SQL over a fixed window and reconcile the result to the fixture.

Common mistakes

Keep one row equal to one durable outcome

  • Emitting rag_retrieval_evaluated before rag evaluation pipeline knows the final outcome.
  • Mixing several grains in one table, which makes counts and rates ambiguous.
  • Replacing controlled categories with raw URLs, payloads, prompts, or error text.
  • Changing a field type in place after saved queries and dashboards depend on it.
  • Adding identifiers without a documented investigation, access, and retention need.

Use the contract

Query and operationalize the event

Related contracts

Send a fixture before production traffic

Create a free API key, send the synthetic event, and inspect the inferred table before connecting a live workflow.

Test this schema