Arize Phoenix Integration¶
Observe with Phoenix, detect with Pisama.
Arize Phoenix (9,200+ stars) is the most popular open-source LLM observability platform. It captures traces from every major agent framework. Pisama adds multi-agent failure detection on top of those traces.
Phoenix tells you what happened. Pisama tells you what went wrong and why.
How It Works¶
Pisama ingests Phoenix trace exports for failure analysis via the built-in PhoenixImporter. (Exporting Pisama results back into Phoenix as spans is on the roadmap, not yet shipped.)
Import Phoenix Traces¶
Option 1: Export from Phoenix UI¶
- In Phoenix, select the traces you want to analyze
- Export as JSON (OTEL format)
- Import into Pisama:
# CLI
pisama analyze phoenix_export.json
# Or via API
curl -X POST https://api.pisama.ai/api/v1/traces \
-H "Authorization: Bearer $PISAMA_API_KEY" \
-H "Content-Type: application/json" \
-d @phoenix_export.json
Option 2: Bulk Import¶
For historical analysis of many traces:
curl -X POST https://api.pisama.ai/api/v1/import-jobs \
-H "Authorization: Bearer $PISAMA_API_KEY" \
-F "file=@phoenix_traces.json" \
-F "format=phoenix"
Monitor import progress:
Option 3: Python SDK¶
from pisama import analyze
# Analyze a Phoenix JSON export directly
result = analyze("phoenix_export.json")
for issue in result.issues:
print(f"[{issue.type}] {issue.summary} (severity: {issue.severity})")
print(f" Agent: {issue.agent_id}")
print(f" {issue.recommendation}")
Export Results to Phoenix¶
Roadmap
Exporting Pisama detection results back into Phoenix as OTEL spans is planned but not yet available. For now, review detections in the Pisama dashboard or via the REST API.
Framework Auto-Detection¶
When importing Phoenix traces, Pisama automatically detects the agent framework from OTEL gen_ai.system and gen_ai.provider.name semantic conventions:
| Framework | Auto-Detected | Additional Detectors |
|---|---|---|
| LangGraph | Yes | 6 LangGraph-specific (recursion, state corruption, etc.) |
| CrewAI | Yes | Core 32 detectors |
| AutoGen | Yes | Core 32 detectors |
| OpenAI Agents SDK | Yes | Core 32 detectors |
| Claude Agent SDK | Yes | Core 32 detectors |
| Amazon Bedrock | Yes | Core 32 detectors |
| Google Vertex AI | Yes | Core 32 detectors |
Typical Workflow¶
- Instrument your agents with Phoenix (one-line auto-instrumentation)
- Run your agents — Phoenix captures all traces
- Export traces from Phoenix (JSON or via API)
- Analyze with Pisama — get failure classifications, severity scores, and fix recommendations
- Review detection results in the Pisama dashboard or via the REST API
Why Both?¶
| Capability | Phoenix | Pisama |
|---|---|---|
| Trace capture & visualization | Yes | Via import |
| LLM call cost tracking | Yes | Yes |
| Multi-agent trace graphs | Yes | Yes |
| Failure pattern detection (84 detectors) | No | Yes |
| Framework-specific detectors (5 platforms) | No | Yes |
| Tiered detection ($0 for 90%+) | No | Yes |
| Self-healing & fix generation | No | Yes |
| F1-calibrated accuracy scores | No | Yes |
Phoenix is the best open-source observability platform. Pisama is the best open-source failure detection platform. Use both.
Supported Formats¶
The Phoenix importer handles:
- Phoenix JSON export: Native format from Phoenix UI export
- OTEL JSON (
resourceSpansformat): Standard OTEL trace export - JSONL: Line-delimited JSON traces
All formats are auto-detected — just pass the file and Pisama figures out the format.