feedback

Log Feedback on an MLflow Trace

Run ${CLAUDE_SKILL_DIR}/scripts/trace_annotate.py to log human feedback (assessments) on traces.

Examples

Positive feedback on the most recent trace:

uv run python ${CLAUDE_SKILL_DIR}/scripts/trace_annotate.py feedback --name thumbs_up --value true

Quality rating with rationale:

uv run python ${CLAUDE_SKILL_DIR}/scripts/trace_annotate.py feedback --name quality --value good --rationale "great answer"

Feedback on a specific trace:

uv run python ${CLAUDE_SKILL_DIR}/scripts/trace_annotate.py feedback --trace-id tr-abc123 --name quality --value poor --rationale "hallucinated the API endpoint"

List recent traces (to find a trace ID):

uv run python ${CLAUDE_SKILL_DIR}/scripts/trace_annotate.py list

Arguments

feedback subcommand

Arg	Required	Description
--name	Yes	Feedback name (e.g. quality, thumbs_up, relevance)
--value	Yes	Feedback value (e.g. good, poor, true, 5)
--trace-id	No	Trace ID (default: most recent trace)
--rationale	No	Free-text explanation of the feedback

list subcommand

Arg	Required	Description
--max-results	No	Number of traces to show (default: 10)

Notes

• Feedback is stored as assessments on the trace's info.assessments array
• A has_feedback=true tag is automatically set for searchability
• Search for traces with feedback: mlflow traces search --filter-string "tag.has_feedback = 'true'"
• Multiple feedback entries can be added to the same trace (e.g., both quality and relevance)
• The source_type='HUMAN' indicates this feedback came from a human reviewer via Claude Code
• Common feedback patterns:
  • thumbs_up / thumbs_down for quick binary feedback
  • quality:good / quality:poor for quality assessment
  • rating:1-5 for numeric scales