Minimum score
score: factuality: 1-5 usefulness: 1-5 format: 1-5 safety: pass|fail needs_human: true|false reason: "invented citation in the second paragraph"
Who scores
- Deterministic rules: valid JSON, complete fields, citations present.
- Automatic evals: compare against expected cases.
- LLM-as-judge: useful for triage, not as absolute truth.
- Human: required for editorial, legal, financial, or reputational content.
Official source
Langfuse documents traces, evaluation, datasets, and scores for self-hostable LLM applications: https://langfuse.com/docs/observability/overview