Phoenix Arize

Phoenix is Arize AI's open-source LLM observability platform — OTel-based tracing with a purpose-built UI for LLM span inspection (token counts, latency, prompts/completions), built-in evaluation pipelines (hallucination, relevance, toxicity), and dataset management for model improvement. Strong for teams that want both traces and evals in the same open-source tool. Self-hostable locally or on Arize Phoenix Cloud. Confidence is docs-derived.

Python SDK (arize-phoenix, openinference-instrumentation-*). One-line auto-instrumentation: OpenAIInstrumentor().instrument(). Traces sent to Phoenix server via OTel OTLP. Local server: import phoenix as px; px.launch_app(). Eval SDK: px.evals for running evaluation suites on trace data. Datasets: collect spans as datasets for fine-tuning or testing. Phoenix Cloud REST API for programmatic trace/dataset access.

Open-source local: no API key required. Phoenix Cloud: API key for sending traces to hosted service. OTel-compatible exporters — trace data travels via standard OTLP. LLM provider keys managed by application, not Phoenix. HTTPS enforced for Phoenix Cloud.

Local server: runs in-process or as a separate server; restart loses in-memory state (use persistence mode for durable traces). Phoenix Cloud: data persistent, uptime tracked. Auto-instrumentation patches: version pinning recommended to avoid instrumentation conflicts. OpenInference semantic conventions for LLM spans are stable.

docs.arize.com/phoenix covers quickstart, tracing guides, eval workflows, dataset management, and Phoenix Cloud setup. Getting started: pip install arize-phoenix openinference-instrumentation-openai, launch local UI, first trace in under 5 minutes. Open-source (GitHub: Arize-ai/phoenix). Community via Arize Slack and GitHub.