OpenInference

OpenInference is a set of semantic conventions and instrumentation plugins that extend OpenTelemetry for tracing AI applications. The specification describes LLM invocations and the surrounding application context, such as retrieval from vector stores and calls to external tools, in a transport-agnostic and file-format-agnostic way so traces can be expressed in JSON, Protobuf, or DataFrames. Arize Phoenix and Arize AX consume these traces natively, and any OpenTelemetry-compatible backend can receive them.

The instrumentation libraries wrap machine-learning SDKs and agent frameworks so their calls emit OpenInference spans automatically. Python has the widest coverage; JavaScript, Java, and Go packages follow the same conventions. Span processors also convert traces produced by OpenLIT and OpenLLMetry into OpenInference format.

Features

• Python instrumentors: OpenAI, OpenAI Agents SDK, Anthropic, Claude Agent SDK, Google GenAI, Google ADK, Vertex AI, AWS Bedrock, Mistral AI, Groq, Together, Ollama, Cohere, and LiteLLM
• Python agent frameworks: LangChain, LlamaIndex, DSPy, CrewAI, Haystack, Agno, AG2, AutoGen AgentChat, Pydantic AI, smolagents, BeeAI, Strands Agents, Pipecat, Instructor, Guardrails, Portkey, and MCP
• JavaScript packages: @arizeai/openinference-core plus instrumentation for OpenAI, Anthropic, Claude Agent SDK, LangChain, Bedrock, Bedrock Agent Runtime, BeeAI, MCP, Vercel AI SDK, TanStack AI, and GenAI
• Java packages: LangChain4j, Spring AI, and an annotation-based instrumentation
• Go packages: OpenAI Go and Anthropic Go SDK instrumentation
• Semantic conventions: shared attribute definitions published per language for hand-written spans
• Span processors: converters for OpenLIT and OpenLLMetry traces
• Backends: Arize Phoenix, Arize AX, or any OpenTelemetry collector