Self-hosted observability, OpenTelemetry native, embedding visualization.

Arize Phoenix — Open-Source LLM Tracing and Evaluation

Arize Phoenix is an open-source, self-hostable LLM observability platform with a focus on trace analysis and embedding visualization. Phoenix is OpenTelemetry-native — uses OTel spans for traces. Works with any LLM framework: LangChain, LlamaIndex, or raw API calls via the openinference SDK. Key differentiator: dimensionality reduction visualization of embeddings to detect drift and clustering issues.

Interactive Pipeline Flow

UnderstandImplementOptimise

🔭 OTel — native tracing

Phoenix is OpenTelemetry-native: traces are OTel spans with LLM semantic conventions. Integrates with standard OTel collector pipelines.

Architecture Diagrams

Instrumenting a Python LLM application with Phoenix.

Loading diagram...

Key Design Decisions

🖥️Phoenix is the go-to for compliance-sensitive self-hosted LLM observability

LangSmith sends traces to LangChain's cloud. Phoenix runs entirely on your infrastructure. For healthcare, finance, or government applications where trace data contains sensitive information, self-hosting is mandatory.

🗺️Embedding visualization reveals retrieval quality visually

UMAP plots of query and document embeddings reveal: are queries clustering far from relevant documents (retrieval quality issue)? Are there topic clusters you haven't optimised prompts for? Visual inspection catches subtle retrieval problems that metrics miss.

🔭OTel-native means Phoenix integrates with your existing observability stack

If you already use Jaeger, Tempo, or Azure Monitor, Phoenix traces can be exported via OTLP. One tracing setup captures both application traces (HTTP, DB) and LLM traces in the same system.

Discussion