otel-eval-bridge

otel-eval-bridge converts exported OpenTelemetry or Phoenix-style GenAI spans into local eval cases while preserving trace and span references.

At a Glance

Install

python -m pip install "otel-eval-bridge==0.1.2"

Verified Quickstart

Run from a source checkout:

otel-eval-bridge extract examples/genai_trace.json \
  --out /tmp/eval-cases.jsonl

otel-eval-bridge manifest /tmp/eval-cases.jsonl \
  --out /tmp/eval-manifest.json

otel-eval-bridge attach /tmp/eval-cases.jsonl examples/eval_results.json \
  --out /tmp/trace-linked-results.json

The bundled trace export produces two eval cases and an auraone.eval-run-manifest.v1 manifest.

Input and Output Contract

The reader accepts a JSON list, a top-level spans list, or OTLP JSON resourceSpans[].scopeSpans[].spans[]. Phoenix-style input.value, output.value, llm.provider, and llm.model_name attributes are normalized to the bridge's GenAI fields.

Each emitted case can include:

• trace ID, span ID, and span name;
• prompt/input, expected value, and observed completion;
• provider and model metadata.

attach matches result entries by trace ID first and span ID second.

Runtime, Data, and Network Boundary

• All commands read and write local JSON or JSONL files.
• The CLI does not receive OTLP traffic, export telemetry, call an observability backend, run an eval, or make network requests.
• Redaction is enabled by default for sensitive keys, sk-..., pypi-..., bearer tokens, and email addresses.
• --no-redact disables that protection. Default redaction is pattern-based and is not a comprehensive privacy or compliance control.

Limitations

• The bridge extracts eval inputs from exported trace data only. It does not run a scorer, validate observability semantics, or prove trace completeness.
• Redaction is pattern-based and intentionally small. Review emitted cases before sharing them outside the originating environment.

Compatibility

The published auraone-agent-studio-open CLI declares otel-eval-bridge>=0.1.0 as a runtime dependency and exposes local trace import and eval-case extraction paths. The JSONL output can also be adapted by an in-house eval runner.

Publication Status

Verified on 2026-07-13:

• PyPI: otel-eval-bridge==0.1.2
• GitHub release: v0.1.2
• The source, wheel, source distribution, and release tag are aligned on 0.1.2.

Next Action

Export a sanitized trace sample, inspect the generated cases for residual sensitive data, then connect the JSONL file to the eval runner that owns scoring.