OpenInference OTEL Tracing

This documentation provides a guide on using OpenInference OTEL tracing decorators and methods for instrumenting functions, chains, agents, and tools using OpenTelemetry. These tools can be combined with, or used in place of, OpenTelemetry instrumentation code. They are designed to simplify the instrumentation process. Let your coding agent set this up. Works with Claude Code, Codex, Cursor, and OpenCode.

Start Phoenix

In one terminal, start Phoenix and leave it running. It serves at http://localhost:6006:

uvx arize-phoenix serve

Already have a Phoenix deployment? Skip this step. px setup can point at any running Phoenix, including a self-hosted one.

Add tracing with a coding agent

In a new terminal, from your app’s root directory, run one of these:

npx -y @arizeai/phoenix-cli setup

npm i -g @arizeai/phoenix-cli && px setup

px setup hands the instrumentation to your coding agent and waits until a real trace arrives. It can also install the Phoenix skills so your agent can query what you capture. See Connect a project with px setup for the full flow, CI usage, and unsupported agents.

Installation

Ensure you have Phoenix OTEL installed:

Python
TypeScript
Go

pip install "arize-phoenix-otel>=0.16.0"

Version arize-phoenix-otel>=0.16.0 is required for the examples on this page. Starting in 0.16.0, phoenix.otel re-exports the OpenInference context managers (using_session, using_user, using_metadata, using_tags, using_attributes, using_prompt_template, suppress_tracing) and semantic conventions (SpanAttributes, OpenInferenceSpanKindValues, OpenInferenceMimeTypeValues) directly — no need to also install openinference-instrumentation or openinference-semantic-conventions.On older versions, import the same symbols from openinference.instrumentation and openinference.semconv.trace instead.

npm install @arizeai/phoenix-otel

@arizeai/phoenix-otel re-exports the OpenInference tracing helpers and context utilities, so Phoenix users can register tracing and add decorators or span helpers from the same package.For detailed API documentation, consult the respective documentation sites.

Phoenix TypeScript API

OpenInference JavaScript API

Phoenix does not ship a phoenix-otel-go wrapper. Install the OpenInference Go semconv + instrumentation packages plus the stock OpenTelemetry Go SDK:

go get github.com/Arize-ai/openinference/go/openinference-semantic-conventions
go get github.com/Arize-ai/openinference/go/openinference-instrumentation
go get go.opentelemetry.io/otel
go get go.opentelemetry.io/otel/exporters/otlp/otlptrace/otlptracehttp
go get go.opentelemetry.io/otel/sdk

The semconv package provides typed constants for the wire-format-stable attribute keys; instrumentation provides context-attribute helpers (session, user, metadata, tags, suppression).

Setting Up Tracing

Python
TypeScript
Go

from phoenix.otel import register

tracer_provider = register(protocol="http/protobuf", project_name="your project name")
tracer = tracer_provider.get_tracer(__name__)

import { register } from "@arizeai/phoenix-otel";

const tracerProvider = register({
  projectName: "my-app",
  url: "https://your-phoenix.com",
  apiKey: process.env.PHOENIX_API_KEY,
});

Configure the OpenTelemetry Go SDK with an OTLP/HTTP exporter pointed at Phoenix, then obtain a tracer. For a remote deployment, set PHOENIX_COLLECTOR_ENDPOINT and, if it has authentication enabled, PHOENIX_API_KEY.

import (
    "context"
    "os"
    "strings"

    "go.opentelemetry.io/otel"
    "go.opentelemetry.io/otel/exporters/otlp/otlptrace/otlptracehttp"
    "go.opentelemetry.io/otel/sdk/resource"
    sdktrace "go.opentelemetry.io/otel/sdk/trace"
    semconvotel "go.opentelemetry.io/otel/semconv/v1.26.0"
)

func phoenixTraceEndpoint(endpoint string) string {
    endpoint = strings.TrimRight(endpoint, "/")
    if strings.HasSuffix(endpoint, "/v1/traces") {
        return endpoint
    }
    return endpoint + "/v1/traces"
}

ctx := context.Background()
opts := []otlptracehttp.Option{
    otlptracehttp.WithEndpoint("localhost:6006"),
    otlptracehttp.WithURLPath("/v1/traces"),
    otlptracehttp.WithInsecure(),
}
if endpoint := os.Getenv("PHOENIX_COLLECTOR_ENDPOINT"); endpoint != "" {
    opts = []otlptracehttp.Option{
        otlptracehttp.WithEndpointURL(phoenixTraceEndpoint(endpoint)),
        otlptracehttp.WithHeaders(map[string]string{
            "Authorization": "Bearer " + os.Getenv("PHOENIX_API_KEY"),
        }),
    }
}

exporter, _ := otlptracehttp.New(ctx, opts...)
res, _ := resource.New(ctx, resource.WithAttributes(
    semconvotel.ServiceName("your project name"),
))
tp := sdktrace.NewTracerProvider(
    sdktrace.WithBatcher(exporter),
    sdktrace.WithResource(res),
)
otel.SetTracerProvider(tp)

tracer := otel.Tracer("your project name")

Using Helpers

Your tracer object can now be used in two primary ways:

1. Tracing a function

Python
TypeScript
Go

@tracer.chain
def my_func(input: str) -> str:
    return "output"

import { traceChain } from "@arizeai/phoenix-otel";

const myFunc = (input: string): string => {
  return "output";
};

const tracedFunc = traceChain(myFunc, { name: "my-func" });

tracedFunc("input");

Go has no decorator syntax. Wrap the function body in a tracer.Start span and set input / output / span-kind attributes manually using the OpenInference Go semconv constants:

import (
    "context"

    "go.opentelemetry.io/otel"
    "go.opentelemetry.io/otel/attribute"
    "go.opentelemetry.io/otel/codes"
    "go.opentelemetry.io/otel/trace"

    "github.com/Arize-ai/openinference/go/openinference-semantic-conventions"
)

func myFunc(ctx context.Context, input string) (string, error) {
    tracer := otel.Tracer("my-app")
    ctx, span := tracer.Start(ctx, "my-func",
        trace.WithAttributes(
            attribute.String(semconv.OpenInferenceSpanKind, semconv.SpanKindChain),
            attribute.String(semconv.InputValue, input),
        ),
    )
    defer span.End()

    output := "output"
    span.SetAttributes(attribute.String(semconv.OutputValue, output))
    span.SetStatus(codes.Ok, "")
    return output, nil
}

This entire function will appear as a Span in Phoenix. Input and output attributes in Phoenix will be set automatically based on my_func’s parameters and return. The status attribute will also be set automatically.

2. As a with clause to trace specific code blocks

Python
TypeScript
Go

with tracer.start_as_current_span(
    "my-span-name",
    openinference_span_kind="chain",
) as span:
    span.set_input("input")
    span.set_output("output")
    span.set_status(Status(StatusCode.OK))

import { withSpan } from "@arizeai/phoenix-otel";
import { trace } from "@arizeai/phoenix-otel";

await withSpan(
  async () => {
    const span = trace.getActiveSpan();
    if (span) {
      span.setAttributes({
        "input.value": "input",
        "output.value": "output",
      });
    }
  },
  {
    name: "my-span-name",
    kind: "CHAIN",
  }
);

Go does not have a with statement; use tracer.Start + defer span.End() to scope a span to a block:

import (
    "go.opentelemetry.io/otel"
    "go.opentelemetry.io/otel/attribute"
    "go.opentelemetry.io/otel/codes"
    "go.opentelemetry.io/otel/trace"

    "github.com/Arize-ai/openinference/go/openinference-semantic-conventions"
)

tracer := otel.Tracer("my-app")
ctx, span := tracer.Start(ctx, "my-span-name",
    trace.WithAttributes(
        attribute.String(semconv.OpenInferenceSpanKind, semconv.SpanKindChain),
    ),
)
defer span.End()

span.SetAttributes(
    attribute.String(semconv.InputValue, "input"),
    attribute.String(semconv.OutputValue, "output"),
)
span.SetStatus(codes.Ok, "")

The code within this clause will be captured as a Span in Phoenix. Here the input, output, and status must be set manually. This approach is useful when you need only a portion of a method to be captured as a Span.

OpenInference Span Kinds

OpenInference Span Kinds denote the possible types of spans you might capture, and will be rendered different in the Phoenix UI. The openinference.span.kind attribute is required for all OpenInference spans and identifies the type of operation being traced. The span kind provides a hint to the tracing backend as to how the trace should be assembled. Valid values include:

Span Kind	Description
LLM	A span that represents a call to a Large Language Model (LLM). For example, an LLM span could be used to represent a call to OpenAI or Llama for chat completions or text generation.
EMBEDDING	A span that represents a call to an LLM or embedding service for generating embeddings. For example, an Embedding span could be used to represent a call to OpenAI to get an ada embedding for retrieval.
CHAIN	A span that represents a starting point or a link between different LLM application steps. For example, a Chain span could be used to represent the beginning of a request to an LLM application or the glue code that passes context from a retriever to an LLM call.
RETRIEVER	A span that represents a data retrieval step. For example, a Retriever span could be used to represent a call to a vector store or a database to fetch documents or information.
RERANKER	A span that represents the reranking of a set of input documents. For example, a cross-encoder may be used to compute the input documents’ relevance scores with respect to a user query, and the top K documents with the highest scores are then returned by the Reranker.
TOOL	A span that represents a call to an external tool such as a calculator, weather API, or any function execution that is invoked by an LLM or agent.
AGENT	A span that encompasses calls to LLMs and Tools. An agent describes a reasoning block that acts on tools using the guidance of an LLM.
GUARDRAIL	A span that represents calls to a component to protect against jailbreak user input prompts by taking action to modify or reject an LLM’s response if it contains undesirable content. For example, a Guardrail span could involve checking if an LLM’s output response contains inappropriate language, via a custom or external guardrail library, and then amending the LLM response to remove references to the inappropriate language.
EVALUATOR	A span that represents a call to a function or process performing an evaluation of the language model’s outputs. Examples include assessing the relevance, correctness, or helpfulness of the language model’s answers.
DECISION	Experimental. A span that represents a call to a decision model, which scores or selects among candidate options provided in the request rather than generating free-form text. Support is still being built out and no decision-model provider is fully supported yet.

Chains

Python
TypeScript
Go

Using Context Managers

with tracer.start_as_current_span(
    "chain-span-with-plain-text-io",
    openinference_span_kind="chain",
) as span:
    span.set_input("input")
    span.set_output("output")
    span.set_status(Status(StatusCode.OK))

Using Decorators

@tracer.chain
def decorated_chain_with_plain_text_output(input: str) -> str:
    return "output"

decorated_chain_with_plain_text_output("input")

Using JSON Output

@tracer.chain
def decorated_chain_with_json_output(input: str) -> Dict[str, Any]:
    return {"output": "output"}

decorated_chain_with_json_output("input")

Overriding Span Name

@tracer.chain(name="decorated-chain-with-overridden-name")
def this_name_should_be_overridden(input: str) -> Dict[str, Any]:
    return {"output": "output"}

this_name_should_be_overridden("input")

Using Wrappers

import { withSpan } from "@arizeai/phoenix-otel";
import { trace } from "@arizeai/phoenix-otel";

await withSpan(
  async () => {
    const span = trace.getActiveSpan();
    if (span) {
      span.setAttributes({
        "input.value": "input",
        "output.value": "output",
      });
    }
  },
  {
    name: "chain-span-with-plain-text-io",
    kind: "CHAIN",
  }
);

Using Function Wrappers

import { traceChain } from "@arizeai/phoenix-otel";

const decoratedChainWithPlainTextOutput = traceChain(
  (input: string): string => {
    return "output";
  },
  { name: "decorated-chain-with-plain-text-output" }
);

decoratedChainWithPlainTextOutput("input");

Using JSON Serializable Output

import { traceChain } from "@arizeai/phoenix-otel";

const decoratedChainWithJsonOutput = traceChain(
  (input: string): Record<string, any> => {
    return { output: "output" };
  },
  { name: "decorated-chain-with-json-output" }
);

decoratedChainWithJsonOutput("input");

Overriding Span Name

import { traceChain } from "@arizeai/phoenix-otel";

const thisNameShouldBeOverridden = traceChain(
  (input: string): Record<string, any> => {
    return { output: "output" };
  },
  { name: "decorated-chain-with-overridden-name" }
);

thisNameShouldBeOverridden("input");

Set the OpenInference span kind to CHAIN on any span that represents a step or glue point in your application. Inputs and outputs go on the same span via the semconv constants:

import (
    "go.opentelemetry.io/otel"
    "go.opentelemetry.io/otel/attribute"
    "go.opentelemetry.io/otel/trace"

    "github.com/Arize-ai/openinference/go/openinference-semantic-conventions"
)

tracer := otel.Tracer("my-app")
ctx, span := tracer.Start(ctx, "my-chain",
    trace.WithAttributes(
        attribute.String(semconv.OpenInferenceSpanKind, semconv.SpanKindChain),
        attribute.String(semconv.InputValue, "input"),
    ),
)
defer span.End()

// ... do work ...

span.SetAttributes(attribute.String(semconv.OutputValue, "output"))

Agents

Python
TypeScript
Go

Using Context Managers

with tracer.start_as_current_span(
    "agent-span-with-plain-text-io",
    openinference_span_kind="agent",
) as span:
    span.set_input("input")
    span.set_output("output")
    span.set_status(Status(StatusCode.OK))

Using Decorators

@tracer.agent
def decorated_agent(input: str) -> str:
    return "output"

decorated_agent("input")

Using Function Wrappers

import { withSpan } from "@arizeai/phoenix-otel";
import { trace } from "@arizeai/phoenix-otel";

await withSpan(
  async () => {
    const span = trace.getActiveSpan();
    if (span) {
      span.setAttributes({
        "input.value": "input",
        "output.value": "output",
      });
    }
  },
  {
    name: "agent-span-with-plain-text-io",
    kind: "AGENT",
  }
);

Using Function Wrappers

import { traceAgent } from "@arizeai/phoenix-otel";

const decoratedAgent = traceAgent(
  (input: string): string => {
    return "output";
  },
  { name: "decorated-agent" }
);

decoratedAgent("input");

Use the AGENT span kind for spans that wrap an LLM-driven reasoning loop. Set agent.name to identify the agent across calls:

import (
    "go.opentelemetry.io/otel"
    "go.opentelemetry.io/otel/attribute"
    "go.opentelemetry.io/otel/trace"

    "github.com/Arize-ai/openinference/go/openinference-semantic-conventions"
)

tracer := otel.Tracer("my-app")
ctx, span := tracer.Start(ctx, "my-agent",
    trace.WithAttributes(
        attribute.String(semconv.OpenInferenceSpanKind, semconv.SpanKindAgent),
        attribute.String(semconv.AgentName, "research-agent"),
        attribute.String(semconv.InputValue, userQuery),
    ),
)
defer span.End()

// ... run agent loop, calling tools and LLMs (those produce child spans) ...

span.SetAttributes(attribute.String(semconv.OutputValue, finalAnswer))

Tools

Python
TypeScript
Go

Using Context Managers

with tracer.start_as_current_span(
    "tool-span",
    openinference_span_kind="tool",
) as span:
    span.set_input("input")
    span.set_output("output")
    span.set_tool(
        name="tool-name",
        description="tool-description",
        parameters={"input": "input"},
    )
    span.set_status(Status(StatusCode.OK))

Using Decorators

@tracer.tool
def decorated_tool(input1: str, input2: int) -> None:
    """
    tool-description
    """

decorated_tool("input1", 1)

Overriding Tool Name

@tracer.tool(
    name="decorated-tool-with-overridden-name",
    description="overridden-tool-description",
)
def this_tool_name_should_be_overridden(input1: str, input2: int) -> None:
    """
    this tool description should be overridden
    """

this_tool_name_should_be_overridden("input1", 1)

Using Function Wrappers

import { traceTool } from "@arizeai/phoenix-otel";

/**
 * tool-description
 */
const decoratedTool = traceTool(
  (input1: string, input2: number): void => {
    // Tool implementation
  },
  { name: "decorated-tool" }
);

decoratedTool("input1", 1);

Overriding Tool Name

import { traceTool } from "@arizeai/phoenix-otel";

/**
 * this tool description should be overridden
 */
const thisToolNameShouldBeOverridden = traceTool(
  (input1: string, input2: number): void => {
    // Tool implementation
  },
  {
    name: "decorated-tool-with-overridden-name",
    description: "overridden-tool-description",
  }
);

thisToolNameShouldBeOverridden("input1", 1);

Use the TOOL span kind for spans that wrap calls to external functions, APIs, or any tool the agent invokes. The tool name and inputs are conventionally attached to the same span:

import (
    "encoding/json"

    "go.opentelemetry.io/otel"
    "go.opentelemetry.io/otel/attribute"
    "go.opentelemetry.io/otel/trace"

    "github.com/Arize-ai/openinference/go/openinference-semantic-conventions"
)

tracer := otel.Tracer("my-app")
toolArgs := map[string]any{"city": "Johannesburg"}
toolArgsJSON, _ := json.Marshal(toolArgs)

ctx, span := tracer.Start(ctx, "lookup_weather",
    trace.WithAttributes(
        attribute.String(semconv.OpenInferenceSpanKind, semconv.SpanKindTool),
        attribute.String(semconv.InputValue, string(toolArgsJSON)),
        attribute.String(semconv.InputMimeType, semconv.MimeTypeJSON),
    ),
)
defer span.End()

result := lookupWeather(toolArgs["city"].(string))
span.SetAttributes(attribute.String(semconv.OutputValue, result))

LLMs

Like other span kinds, LLM spans can be instrumented either via a context manager or via a decorator pattern. It’s also possible to directly patch client methods. While this guide uses the OpenAI Python client for illustration, in practice, you should use the OpenInference auto-instrumentors for OpenAI whenever possible and resort to manual instrumentation for LLM spans only as a last resort. To run the snippets in this section, set your OPENAI_API_KEY environment variable.

Python
TypeScript
Go

Context Manager

from openai import OpenAI
from opentelemetry.trace import Status, StatusCode

openai_client = OpenAI()

messages = [{"role": "user", "content": "Hello, world!"}]
with tracer.start_as_current_span("llm_span", openinference_span_kind="llm") as span:
    span.set_input(messages)
    try:
        response = openai_client.chat.completions.create(
            model="gpt-5.6-luna",
            messages=messages,
        )
    except Exception as error:
        span.record_exception(error)
        span.set_status(Status(StatusCode.ERROR))
    else:
        span.set_output(response)
        span.set_status(Status(StatusCode.OK))

Decorator

from typing import List

from openai import OpenAI
from openai.types.chat import ChatCompletionMessageParam

openai_client = OpenAI()


@tracer.llm
def invoke_llm(
    messages: List[ChatCompletionMessageParam],
) -> str:
    response = openai_client.chat.completions.create(
        model="gpt-4o",
        messages=messages,
    )
    message = response.choices[0].message
    return message.content or ""


invoke_llm([{"role": "user", "content": "Hello, world!"}])

This decorator pattern above works for sync functions, async coroutine functions, sync generator functions, and async generator functions. Here’s an example with an async generator.

from typing import AsyncGenerator, List

from openai import AsyncOpenAI
from openai.types.chat import ChatCompletionMessageParam

openai_async_client = AsyncOpenAI()


@tracer.llm
async def stream_llm_responses(
    messages: List[ChatCompletionMessageParam],
) -> AsyncGenerator[str, None]:
    stream = await openai_async_client.chat.completions.create(
        model="gpt-4o",
        messages=messages,
        stream=True,
    )
    async for chunk in stream:
        if chunk.choices[0].delta.content:
            yield chunk.choices[0].delta.content


# invoke inside of an async context
async for token in stream_llm_responses([{"role": "user", "content": "Hello, world!"}]):
    print(token, end="")

Method Patch

It’s also possible to directly patch methods on a client. This is useful if you want to transparently use the client in your application with instrumentation logic localized in one place.

from openai import OpenAI

openai_client = OpenAI()

# patch the create method
wrapper = tracer.llm
openai_client.chat.completions.create = wrapper(openai_client.chat.completions.create)

# invoke the patched method normally
openai_client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Hello, world!"}],
)

The snippets above produce LLM spans with input and output values, but don’t offer rich UI for messages, tools, invocation parameters, etc. In order to manually instrument LLM spans with these features, users can define their own functions to wrangle the input and output of their LLM calls into OpenInference format. These richer LLM attribute helpers are not re-exported from phoenix.otel — install openinference-instrumentation to access them:

get_llm_attributes
get_input_attributes
get_output_attributes

pip install openinference-instrumentation

For OpenAI, these functions might look like this:

from typing import Any, Dict, List, Optional, Union

from openai.types.chat import (
    ChatCompletion,
    ChatCompletionMessage,
    ChatCompletionMessageParam,
    ChatCompletionToolParam,
)
from opentelemetry.util.types import AttributeValue

import openinference.instrumentation as oi
from openinference.instrumentation import (
    get_input_attributes,
    get_llm_attributes,
    get_output_attributes,
)


def process_input(
    messages: List[ChatCompletionMessageParam],
    model: str,
    temperature: Optional[float] = None,
    tools: Optional[List[ChatCompletionToolParam]] = None,
    **kwargs: Any,
) -> Dict[str, AttributeValue]:
    oi_messages = [convert_openai_message_to_oi_message(message) for message in messages]
    oi_tools = [convert_openai_tool_param_to_oi_tool(tool) for tool in tools or []]
    return {
        **get_input_attributes(
            {
                "messages": messages,
                "model": model,
                "temperature": temperature,
                "tools": tools,
                **kwargs,
            }
        ),
        **get_llm_attributes(
            provider="openai",
            system="openai",
            model_name=model,
            input_messages=oi_messages,
            invocation_parameters={"temperature": temperature},
            tools=oi_tools,
        ),
    }


def convert_openai_message_to_oi_message(
    message_param: Union[ChatCompletionMessageParam, ChatCompletionMessage],
) -> oi.Message:
    if isinstance(message_param, ChatCompletionMessage):
        role: str = message_param.role
        oi_message = oi.Message(role=role)
        if isinstance(content := message_param.content, str):
            oi_message["content"] = content
        if message_param.tool_calls is not None:
            oi_tool_calls: List[oi.ToolCall] = []
            for tool_call in message_param.tool_calls:
                function = tool_call.function
                oi_tool_calls.append(
                    oi.ToolCall(
                        id=tool_call.id,
                        function=oi.ToolCallFunction(
                            name=function.name,
                            arguments=function.arguments,
                        ),
                    )
                )
            oi_message["tool_calls"] = oi_tool_calls
        return oi_message

    role = message_param["role"]
    assert isinstance(message_param["content"], str)
    content = message_param["content"]
    return oi.Message(role=role, content=content)


def convert_openai_tool_param_to_oi_tool(tool_param: ChatCompletionToolParam) -> oi.Tool:
    assert tool_param["type"] == "function"
    return oi.Tool(json_schema=dict(tool_param))


def process_output(response: ChatCompletion) -> Dict[str, AttributeValue]:
    message = response.choices[0].message
    role = message.role
    oi_message = oi.Message(role=role)
    if isinstance(message.content, str):
        oi_message["content"] = message.content
    if isinstance(message.tool_calls, list):
        oi_tool_calls: List[oi.ToolCall] = []
        for tool_call in message.tool_calls:
            tool_call_id = tool_call.id
            function_name = tool_call.function.name
            function_arguments = tool_call.function.arguments
            oi_tool_calls.append(
                oi.ToolCall(
                    id=tool_call_id,
                    function=oi.ToolCallFunction(
                        name=function_name,
                        arguments=function_arguments,
                    ),
                )
            )
        oi_message["tool_calls"] = oi_tool_calls
    output_messages = [oi_message]
    token_usage = response.usage
    oi_token_count: Optional[oi.TokenCount] = None
    if token_usage is not None:
        prompt_tokens = token_usage.prompt_tokens
        completion_tokens = token_usage.completion_tokens
        oi_token_count = oi.TokenCount(
            prompt=prompt_tokens,
            completion=completion_tokens,
        )
    return {
        **get_llm_attributes(
            output_messages=output_messages,
            token_count=oi_token_count,
        ),
        **get_output_attributes(response),
    }

Context Manager

When using a context manager to create LLM spans, these functions can be used to wrangle inputs and outputs.

import json

from openai import OpenAI
from openai.types.chat import (
    ChatCompletionMessage,
    ChatCompletionMessageParam,
    ChatCompletionToolMessageParam,
    ChatCompletionToolParam,
    ChatCompletionUserMessageParam,
)
from opentelemetry.trace import Status, StatusCode

openai_client = OpenAI()


@tracer.tool
def get_weather(city: str) -> str:
    # make an call to a weather API here
    return "sunny"


messages: List[Union[ChatCompletionMessage, ChatCompletionMessageParam]] = [
    ChatCompletionUserMessageParam(
        role="user",
        content="What's the weather like in San Francisco?",
    )
]
temperature = 0.5
invocation_parameters = {"temperature": temperature}
tools: List[ChatCompletionToolParam] = [
    {
        "type": "function",
        "function": {
            "name": "get_weather",
            "description": "finds the weather for a given city",
            "parameters": {
                "type": "object",
                "properties": {
                    "city": {
                        "type": "string",
                        "description": "The city to find the weather for, e.g. 'London'",
                    }
                },
                "required": ["city"],
            },
        },
    },
]

with tracer.start_as_current_span(
    "llm_tool_call",
    attributes=process_input(
        messages=messages,
        invocation_parameters={"temperature": temperature},
        model="gpt-5.6-luna",
    ),
    openinference_span_kind="llm",
) as span:
    try:
        response = openai_client.chat.completions.create(
            model="gpt-4o",
            messages=messages,
            temperature=temperature,
            tools=tools,
        )
    except Exception as error:
        span.record_exception(error)
        span.set_status(Status(StatusCode.ERROR))
    else:
        span.set_attributes(process_output(response))
        span.set_status(Status(StatusCode.OK))

output_message = response.choices[0].message
tool_calls = output_message.tool_calls
assert tool_calls and len(tool_calls) == 1
tool_call = tool_calls[0]
city = json.loads(tool_call.function.arguments)["city"]
weather = get_weather(city)
messages.append(output_message)
messages.append(
    ChatCompletionToolMessageParam(
        content=weather,
        role="tool",
        tool_call_id=tool_call.id,
    )
)

with tracer.start_as_current_span(
    "tool_call_response",
    attributes=process_input(
        messages=messages,
        invocation_parameters={"temperature": temperature},
        model="gpt-5.6-luna",
    ),
    openinference_span_kind="llm",
) as span:
    try:
        response = openai_client.chat.completions.create(
            model="gpt-4o",
            messages=messages,
            temperature=temperature,
        )
    except Exception as error:
        span.record_exception(error)
        span.set_status(Status(StatusCode.ERROR))
    else:
        span.set_attributes(process_output(response))
        span.set_status(Status(StatusCode.OK))

Decorator

When using the tracer.llm decorator, these functions are passed via the process_input and process_output parameters and should satisfy the following:

The input signature of process_input should exactly match the input signature of the decorated function.
The input signature of process_output has a single argument, the output of the decorated function. This argument accepts the returned value when the decorated function is a sync or async function, or a list of yielded values when the decorated function is a sync or async generator function.
Both process_input and process_output should output a dictionary mapping attribute names to values.

from openai import NOT_GIVEN, OpenAI
from openai.types.chat import ChatCompletion

openai_client = OpenAI()


@tracer.llm(
    process_input=process_input,
    process_output=process_output,
)
def invoke_llm(
    messages: List[ChatCompletionMessageParam],
    model: str,
    temperature: Optional[float] = None,
    tools: Optional[List[ChatCompletionToolParam]] = None,
) -> ChatCompletion:
    response: ChatCompletion = openai_client.chat.completions.create(
        messages=messages,
        model=model,
        tools=tools or NOT_GIVEN,
        temperature=temperature,
    )
    return response


invoke_llm(
    messages=[{"role": "user", "content": "Hello, world!"}],
    temperature=0.5,
    model="gpt-5.6-luna",
)

When decorating a generator function, process_output should accept a single argument, a list of the values yielded by the decorated function.

from typing import Dict, List, Optional

from openai.types.chat import ChatCompletionChunk
from opentelemetry.util.types import AttributeValue

import openinference.instrumentation as oi
from openinference.instrumentation import (
    get_llm_attributes,
    get_output_attributes,
)


def process_generator_output(
    outputs: List[ChatCompletionChunk],
) -> Dict[str, AttributeValue]:
    role: Optional[str] = None
    content = ""
    oi_token_count = oi.TokenCount()
    for chunk in outputs:
        if choices := chunk.choices:
            assert len(choices) == 1
            delta = choices[0].delta
            if isinstance(delta.content, str):
                content += delta.content
            if isinstance(delta.role, str):
                role = delta.role
        if (usage := chunk.usage) is not None:
            if (prompt_tokens := usage.prompt_tokens) is not None:
                oi_token_count["prompt"] = prompt_tokens
            if (completion_tokens := usage.completion_tokens) is not None:
                oi_token_count["completion"] = completion_tokens
    oi_messages = []
    if role and content:
        oi_messages.append(oi.Message(role=role, content=content))
    return {
        **get_llm_attributes(
            output_messages=oi_messages,
            token_count=oi_token_count,
        ),
        **get_output_attributes(content),
    }

Then the decoration is the same as before.

from typing import AsyncGenerator

from openai import AsyncOpenAI
from openai.types.chat import ChatCompletionChunk

openai_async_client = AsyncOpenAI()


@tracer.llm(
    process_input=process_input,  # same as before
    process_output=process_generator_output,
)
async def stream_llm_response(
    messages: List[ChatCompletionMessageParam],
    model: str,
    temperature: Optional[float] = None,
) -> AsyncGenerator[ChatCompletionChunk, None]:
    async for chunk in await openai_async_client.chat.completions.create(
        messages=messages,
        model=model,
        temperature=temperature,
        stream=True,
    ):
        yield chunk


async for chunk in stream_llm_response(
    messages=[{"role": "user", "content": "Hello, world!"}],
    temperature=0.5,
    model="gpt-5.6-luna",
):
    print(chunk)

Method Patch

As before, it’s possible to directly patch the method on the client. Just ensure that the input signatures of process_input and the patched method match.

from openai import OpenAI
from openai.types.chat import ChatCompletionMessageParam

openai_client = OpenAI()

# patch the create method
wrapper = tracer.llm(
    process_input=process_input,
    process_output=process_output,
)
openai_client.chat.completions.create = wrapper(openai_client.chat.completions.create)

# invoke the patched method normally
openai_client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Hello, world!"}],
)

Function Wrapper

import { withSpan } from "@arizeai/phoenix-otel";
import OpenAI from "openai";

const openaiClient = new OpenAI();

const invokeLLM = withSpan(
  async (
    messages: Array<{ role: string; content: string }>
  ): Promise<string> => {
    const response = await openaiClient.chat.completions.create({
      model: "gpt-4o",
      messages: messages,
    });
    const message = response.choices[0].message;
    return message.content || "";
  },
  {
    name: "invoke-llm",
    kind: "LLM",
  }
);

await invokeLLM([{ role: "user", content: "Hello, world!" }]);

The snippets above produce LLM spans with input and output values, but don’t offer rich UI for messages, tools, invocation parameters, etc. In order to manually instrument LLM spans with these features, users can use helper functions from @arizeai/phoenix-otel that produce valid OpenInference attributes for LLM spans:

getLLMAttributes
defaultProcessInput
defaultProcessOutput

For OpenAI, these functions might look like this:

import {
  getLLMAttributes,
  defaultProcessInput,
  defaultProcessOutput,
} from "@arizeai/phoenix-otel";
import OpenAI from "openai";

interface ChatCompletionMessageParam {
  role: string;
  content: string;
}

interface ChatCompletionToolParam {
  type: string;
  function: {
    name: string;
    description: string;
    parameters: Record<string, unknown>;
  };
}

function processInput(
  messages: ChatCompletionMessageParam[],
  model: string,
  temperature?: number,
  tools?: ChatCompletionToolParam[]
) {
  const inputAttrs = defaultProcessInput({
    messages,
    model,
    temperature,
    tools,
  });
  const llmAttrs = getLLMAttributes({
    provider: "openai",
    system: "openai",
    modelName: model,
    inputMessages: messages.map((msg) => ({
      role: msg.role,
      content: msg.content,
    })),
    invocationParameters: { temperature },
    tools: tools?.map((tool) => ({ jsonSchema: tool })),
  });
  return { ...inputAttrs, ...llmAttrs };
}

function processOutput(response: OpenAI.Chat.Completions.ChatCompletion) {
  const message = response.choices[0].message;
  const outputAttrs = defaultProcessOutput(response);
  const llmAttrs = getLLMAttributes({
    outputMessages: [
      {
        role: message.role,
        content: typeof message.content === "string" ? message.content : "",
      },
    ],
    tokenCount: response.usage
      ? {
          prompt: response.usage.prompt_tokens,
          completion: response.usage.completion_tokens,
          total: response.usage.total_tokens,
        }
      : undefined,
  });
  return { ...outputAttrs, ...llmAttrs };
}

Function Wrapper

When using withSpan to wrap functions, you can pass processInput and processOutput functions as options. These should satisfy the following:

The input signature of processInput should exactly match the input signature of the wrapped function.
The input signature of processOutput has a single argument, the output of the wrapped function. This argument accepts the returned value when the wrapped function is a sync or async function.
Both processInput and processOutput should output a dictionary mapping attribute names to values.

import { withSpan } from "@arizeai/phoenix-otel";
import OpenAI from "openai";

const openaiClient = new OpenAI();

const invokeLLM = withSpan(
  async (
    messages: ChatCompletionMessageParam[],
    model: string,
    temperature?: number,
    tools?: ChatCompletionToolParam[]
  ): Promise<OpenAI.Chat.Completions.ChatCompletion> => {
    const response = await openaiClient.chat.completions.create({
      messages: messages,
      model: model,
      tools: tools,
      temperature: temperature,
    });
    return response;
  },
  {
    name: "invoke-llm",
    kind: "LLM",
    processInput: (messages, model, temperature, tools) =>
      processInput(messages, model, temperature, tools),
    processOutput: (response) => processOutput(response),
  }
);

await invokeLLM([{ role: "user", content: "Hello, world!" }], "gpt-5.6-luna", 0.5);

For SDKs with OpenInference Go auto-instrumentation, see OpenAI Go SDK and Anthropic SDK Go — no manual span code required. For SDKs without auto-instrumentation, attach model and token attributes directly using semconv:

import (
    "go.opentelemetry.io/otel"
    "go.opentelemetry.io/otel/attribute"
    "go.opentelemetry.io/otel/trace"

    "github.com/Arize-ai/openinference/go/openinference-semantic-conventions"
)

tracer := otel.Tracer("my-app")
ctx, span := tracer.Start(ctx, "invoke_llm",
    trace.WithAttributes(
        attribute.String(semconv.OpenInferenceSpanKind, semconv.SpanKindLLM),
        attribute.String(semconv.LLMProvider, semconv.LLMProviderOpenAI),
        attribute.String(semconv.LLMSystem, semconv.LLMSystemOpenAI),
        attribute.String(semconv.LLMModelName, "gpt-5.6-luna"),
        attribute.String(semconv.InputValue, "Hello, world!"),
    ),
)
defer span.End()

// ... make the LLM call ...

span.SetAttributes(
    attribute.String(semconv.OutputValue, completionText),
    attribute.Int(semconv.LLMTokenCountPrompt, promptTokens),
    attribute.Int(semconv.LLMTokenCountCompletion, completionTokens),
    attribute.Int(semconv.LLMTokenCountTotal, totalTokens),
)

Additional Features

The OpenInference Tracer shown above respects context Managers for Suppressing Tracing & Adding Metadata

Suppress Tracing

Python
TypeScript
Go

from phoenix.otel import suppress_tracing

with suppress_tracing():
    # this trace will not be recorded
    with tracer.start_as_current_span(
        "THIS-SPAN-SHOULD-NOT-BE-TRACED",
        openinference_span_kind="chain",
    ) as span:
        span.set_input("input")
        span.set_output("output")
        span.set_status(Status(StatusCode.OK))

import { withSpan } from "@arizeai/phoenix-otel";
import { trace, context, suppressTracing } from "@arizeai/phoenix-otel";

await context.with(suppressTracing(context.active()), async () => {
  // this trace will not be recorded
  await withSpan(
    async () => {
      const span = trace.getActiveSpan();
      if (span) {
        span.setAttributes({
          "input.value": "input",
          "output.value": "output",
        });
      }
    },
    {
      name: "THIS-SPAN-SHOULD-NOT-BE-TRACED",
      kind: "CHAIN",
    }
  );
});

Use instrumentation.WithSuppression to mark a context such that any OpenInference-aware Go instrumentor (e.g. the OpenAI and Anthropic Go transports / middlewares) skips span emission for calls descended from it. Typical use: evaluator or grader code that itself calls an LLM but should not appear in the customer’s product trace.

import (
    "github.com/Arize-ai/openinference/go/openinference-instrumentation"
)

ctx := instrumentation.WithSuppression(ctx)
resp, _ := client.CreateChatCompletion(ctx, req)  // no LLM span emitted

For manual spans you author yourself, check the context before calling tracer.Start if you want to honor suppression in your own code.

Using Context Attributes

Python
TypeScript
Go

from phoenix.otel import using_attributes

with using_attributes(session_id="123"):
    # this trace has session id "123"
    with tracer.start_as_current_span(
        "chain-span-with-context-attributes",
        openinference_span_kind="chain",
    ) as span:
        span.set_input("input")
        span.set_output("output")
        span.set_status(Status(StatusCode.OK))

import { context } from "@opentelemetry/api";
import { setSession } from "@arizeai/phoenix-otel";
import { withSpan } from "@arizeai/phoenix-otel";
import { trace } from "@arizeai/phoenix-otel";

await context.with(
  setSession(context.active(), { sessionId: "123" }),
  async () => {
    // this trace has session id "123"
    await withSpan(
      async () => {
        const span = trace.getActiveSpan();
        if (span) {
          span.setAttributes({
            "input.value": "input",
            "output.value": "output",
          });
        }
      },
      {
        name: "chain-span-with-context-attributes",
        kind: "CHAIN",
      }
    );
  }
);

Use the helpers in github.com/Arize-ai/openinference/go/openinference-instrumentation to store session, user, metadata, and tags on context.Context. Auto-instrumentors (OpenAI Go, Anthropic Go) automatically apply these to every LLM span descended from the context:

import (
    "encoding/json"

    "github.com/Arize-ai/openinference/go/openinference-instrumentation"
)

metadata := map[string]any{"key-1": "value_1", "key-2": "value_2"}
metadataJSON, _ := json.Marshal(metadata)

ctx = instrumentation.WithSession(ctx, "my-session-id")
ctx = instrumentation.WithUser(ctx, "my-user-id")
ctx = instrumentation.WithMetadata(ctx, string(metadataJSON))
ctx = instrumentation.WithTags(ctx, "tag_1", "tag_2")

// All LLM spans descended from ctx now carry session.id, user.id,
// metadata, and tag.tags — no per-span code needed.
resp, _ := client.CreateChatCompletion(ctx, req)

For manual spans you author yourself, set the same attributes directly via the semconv keys (SessionID, UserID, Metadata, TagTags) after tracer.Start.

Adding Images to your Traces

OpenInference includes message types that can be useful in composing text and image or other file inputs and outputs:

Python
TypeScript
Go

import openinference.instrumentation as oi

image_url = "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg"
text = "describe the weather in this image"
content = [
        {"type": "text", "text": text},
        {
            "type": "image_url",
            "image_url": {"url": image_url, "detail": "low"},
        },
    ]

image = oi.Image(url=image_url)
contents = [
    oi.TextMessageContent(
        type="text",
        text=text,
    ),
    oi.ImageMessageContent(
        type="image",
        image=image,
    ),
]
messages = [
    oi.Message(
        role="user",
        contents=contents,
    )
]

with tracer.start_as_current_span(
    "my-span-name",
    openinference_span_kind="llm",
    attributes=oi.get_llm_attributes(input_messages=messages)
) as span:
    span.set_input(text)

    # Call your LLM here
    response = "This is a test response"

    span.set_output(response)
    print(response.content)

import { withSpan, getLLMAttributes } from "@arizeai/phoenix-otel";
import { trace } from "@arizeai/phoenix-otel";

const imageUrl =
  "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg";
const text = "describe the weather in this image";

const messages = [
  {
    role: "user",
    contents: [
      {
        type: "text",
        text: text,
      },
      {
        type: "image",
        image: {
          url: imageUrl,
        },
      },
    ],
  },
];

await withSpan(
  async () => {
    const span = trace.getActiveSpan();
    if (span) {
      span.setAttributes(
        getLLMAttributes({
          inputMessages: messages,
        })
      );
      // Call your LLM here
      const response = "This is a test response";

      span.setAttributes({
        "input.value": text,
        "output.value": response,
      });
      console.log(response);
    }
  },
  {
    name: "my-span-name",
    kind: "LLM",
  }
);

For SDKs with OpenInference Go auto-instrumentation (OpenAI Go, Anthropic Go), multimodal content is captured automatically when you call the SDK with image-containing messages — no per-span code required.For manual spans, attach image content using the indexed llm.input_messages.{i}.message.contents.{j}.* attributes from github.com/Arize-ai/openinference/go/openinference-semantic-conventions. The Python openinference.instrumentation message-builder helpers do not yet have a Go equivalent — refer to the OpenInference spec for the full multimodal attribute taxonomy.

​OpenInference OTEL Tracing

​Installation

Phoenix TypeScript API

OpenInference JavaScript API

​Setting Up Tracing

​Using Helpers

​1. Tracing a function

​2. As a with clause to trace specific code blocks

​OpenInference Span Kinds

​Chains

​Using Context Managers

​Using Decorators

​Using Wrappers

​Using Function Wrappers

​Agents

​Using Context Managers

​Using Decorators

​Using Function Wrappers

​Using Function Wrappers

​Tools

​Using Context Managers

​Using Decorators

​Using Function Wrappers

​LLMs

​Context Manager

​Decorator

​Method Patch

​Context Manager

​Decorator

​Method Patch

​Function Wrapper

​Function Wrapper

​Additional Features

​Suppress Tracing

​Using Context Attributes

​Adding Images to your Traces

OpenInference OTEL Tracing

Installation

Setting Up Tracing

Using Helpers

1. Tracing a function

2. As a with clause to trace specific code blocks

OpenInference Span Kinds

Chains

Using Context Managers

Using Decorators

Using Wrappers

Using Function Wrappers

Agents

Using Context Managers

Using Decorators

Using Function Wrappers

Using Function Wrappers

Tools

Using Context Managers

Using Decorators

Using Function Wrappers

LLMs

Context Manager

Decorator

Method Patch

Context Manager

Decorator

Method Patch

Function Wrapper

Function Wrapper

Additional Features

Suppress Tracing

Using Context Attributes

Adding Images to your Traces