Skip to content

Phoenix / OpenInference

Kalibra auto-detects OpenInference trace exports from Phoenix. No field mapping needed — just export your spans and compare.

Open in Colab Try the interactive tutorial — run a live agent with your Anthropic key, or use pre-recorded traces to explore instantly.

The workflow

Phoenix (tracing) → Export spans as JSONL → Kalibra (compare)

1. Export traces from Phoenix

from phoenix.client import Client

client = Client()
spans = client.spans.get_spans(project_identifier="default")

with open("traces.jsonl", "w") as f:
    for span in spans:
        f.write(json.dumps(span, default=str) + "\n")

2. Compare with Kalibra

kalibra compare --baseline before.jsonl --current after.jsonl -v

from kalibra.loader import load_traces
from kalibra.engine import compare
from kalibra.renderers import render

baseline = load_traces("before.jsonl")
current = load_traces("after.jsonl")

result = compare(baseline, current, require=["token_delta_pct <= 50"])
print(render(result, "terminal", verbose=True))

That's it. Kalibra detects the OpenInference format from the first line and handles the rest.

What happens under the hood

OpenInference exports are flat arrays of spans — one span per line, each carrying a context.trace_id. Kalibra:

  1. Groups spans into traces by trace_id
  2. Builds the span tree using parent_id relationships
  3. Extracts tokens and cost from LLM spans (CHAIN/AGENT/TOOL spans have None for these — they're excluded automatically)
  4. Counts steps as leaf spans, not total spans — orchestration wrappers aren't actions
  5. Computes duration as wall-clock time (max(end) - min(start)), not sum of span durations
  6. Parses finish reason from llm.output_messages attributes or the output.value JSON (stop_reason for Anthropic, finish_reason for OpenAI/Google) — used by the error_rate and success_rate metrics to detect truncation and errors

What's detected

Format Auto-detected Notes
Phoenix JSONL export (get_spans()) Yes Nested or dot-flattened attributes
Phoenix JSON array export Yes Same detection logic
Flat Kalibra JSONL (one trace per line) Yes Default format, no OpenInference needed
Nested OTel (child_spans trees) Yes JSON array format