Phoenix V2

Phoenix is the local-first LLM observability tool — launch a local server (px.launch_app()), instrument your LLM application with OTel auto-instrumentation, and get an interactive UI showing traces, spans, token counts, and LLM inputs/outputs without sending data to any cloud service. For agents: debug multi-step agent traces locally during development, run automated evals (hallucination detection, Q&A correctness) on trace datasets, and annotate examples for fine-tuning datasets. Phoenix Cloud provides a hosted option for team sharing. Distinct from Arize AI (the cloud ML monitoring platform). Confidence is docs-derived.

Python SDK: pip install arize-phoenix. Launch local UI: import phoenix as px; session = px.launch_app(); print(session.url). Auto-instrument OpenAI: from openinference.instrumentation.openai import OpenAIInstrumentor; OpenAIInstrumentor().instrument(tracer_provider=px.get_default_tracer_provider()). Auto-instrument LangChain: from openinference.instrumentation.langchain import LangChainInstrumentor; LangChainInstrumentor().instrument(). Auto-instrument LlamaIndex: from openinference.instrumentation.llama_index import LlamaIndexInstrumentor; LlamaIndexInstrumentor().instrument(). Evals: from phoenix.evals import OpenAIModel, HallucinationEvaluator, run_evals; eval_model = OpenAIModel(model="gpt-4o"); hallucination_evaluator = HallucinationEvaluator(eval_model); results = run_evals(dataframe=traces_df, evaluators=[hallucination_evaluator]). Phoenix Cloud export: PHOENIX_API_KEY env var + PHOENIX_COLLECTOR_ENDPOINT → traces sent to hosted Phoenix.

Local mode: no auth — localhost:6006 only. Phoenix Cloud: API key auth — PHOENIX_API_KEY env var. Keys from app.phoenix.arize.com → Settings. HTTPS enforced for Phoenix Cloud. Self-hosted Phoenix server: optional auth configuration. OTel instrumentation: no auth required for instrumentation (pushes to local Phoenix endpoint). Data: local mode keeps all data on-machine; Phoenix Cloud sends traces to Arize infrastructure.

Local mode: in-process Python server — always available during development. Trace buffering: OTel spans buffered and exported in background. Large trace datasets: Phoenix UI loads all traces in memory — filter/sample for large datasets. Eval latency: LLM-as-judge evals (HallucinationEvaluator) make LLM API calls per trace — budget accordingly. Phoenix Cloud: async trace ingestion. Concurrent instrumentation: multiple frameworks (OpenAI + LangChain) can be instrumented simultaneously — traces linked by trace_id. UI features: flamegraph, span details, token breakdown, latency histogram.

docs.arize.com/phoenix covers local quickstart, auto-instrumentation guides, eval library, dataset annotation, and Phoenix Cloud. Getting started: pip install arize-phoenix, px.launch_app() — local observability UI in 2 minutes. Phoenix Cloud: free tier. Local: fully free and open-source. GitHub: Arize-ai/phoenix (4k+ stars). Community via Arize Slack (active). Excellent documentation; eval library reference is the standout — comprehensive evaluation patterns for RAG and LLM quality.