TraceroAI

What a trace contains

What TraceroAI stores with every trace, where that data can go, and how to remove it.

TraceroAI stores what your application sends it, in your own Postgres. This page lists exactly what that is, where it can go, and how to remove it, so you can decide what is safe to send.

Stored with every trace#

PartFieldsNotes
Questionthe user's question, and the rewritten retrieval query if you log oneVerbatim. If your users type personal details into questions, they are stored.
Retrieved contexteach chunk's text, source, rank, scores and metadataThe full text of every chunk you log.
Promptthe prompt text, version and template nameOnly if you call log_prompt.
Answerthe generated answer, model, provider, temperature and other parameters, token counts
Timingtotal latency and per-step timings
Projectproject id, environment, pipeline version
Metadatawhatever you pass as metadata=Free-form. Recovery runs add a run id and attempt number here.

Added by the server#

  • Evaluations and diagnosis: scores, the judge's reasons, and the claims it extracted from the answer (quoted from the answer).
  • Feedback: a thumbs up or down and an optional comment of up to 2,000 characters.
  • Time received.

The server doesn't store the caller's IP address or user agent with a trace. The address is used only for short-lived rate-limit counters (in Redis, or in memory when Redis isn't configured) that expire within minutes.

Where it can go#

  • Your database. Traces live in the traces table of the Postgres you configure.
  • Your LLM provider, only if you configure one. With TRACEROAI_OPENAI_API_KEY (or TRACEROAI_JUDGE_BASE_URL) set, the background evaluation sends the question, the retrieved chunks and the answer to that provider for embeddings and the LLM judge. Without a provider key nothing leaves your servers, and diagnoses are term-overlap estimates. Point TRACEROAI_JUDGE_BASE_URL at a model you host to keep judging in-house.
  • Nowhere else. TraceroAI sends no telemetry of its own.

Keeping less#

  • Don't log what you don't need: log_prompt is optional, and chunk text can be trimmed before you pass it to log_retrieval.
  • Remove personal data from questions and answers in your application before sending them. TraceroAI doesn't redact anything yet.
  • Use project keys so each application reads only its own traces.

Removing it#

  • Automatically: set TRACEROAI_TRACE_RETENTION_DAYS. The worker deletes traces older than that every hour. Without the worker (no Redis), run python -m app.retention from services/api on a schedule, for example a daily cron.
  • One trace: DELETE /v1/traces/{trace_id}. A project key can delete only its own project's traces; an admin key, any.
  • Everything for a project: DELETE FROM traces WHERE project_id = '...' in Postgres.