TruLens

Eval without step-level traces cannot distinguish a bad retrieval from a bad generator or a bad tool call. TruLens is the research pick when Evaluations must attach scores to OTEL spans and compare app versions on quality × latency × cost — complementary to Promptfoo/DeepEval CI matrices and to process mining (pm4py) on XES.

1. pip install trulens (+ provider/app extras).
2. @instrument spans (retrieval, generation, tools, MCP, …).
3. Attach Metrics/feedback functions (inline recorder or batch RunConfig).
4. Compare versions on a leaderboard; export OTLP elsewhere if needed.

Related: abm-agent-behavior-mining · agent-evaluation · deepeval · topics/32-eval-harnesses

• Use when: you need tracing + LLM judges on agent/RAG steps, version compare, or OTEL interoperability.
• Skip when: you only need declarative YAML red-team CLI (promptfoo), pytest metric suites without tracing (deepeval), or XES process mining (pm4py).
• Prefer instead: Langfuse/Phoenix-class tools if you already standardized another OTEL UI; keep TruLens when feedback functions + agent GPA-style evals matter.

• Package sprawl: providers/apps are split — missing extras silently disable features.
• Judge calibration: marketed human-agreement figures are context-specific; re-validate on your goldens.
• Not the ledger: traces/scores are evidence, not declared authority or money movement.
• Overlap with observability cell: clarify whether TruLens is eval or ops UI in the install to avoid two systems of dashboards.