Using Tracing Helpers

OpenInference packages provide helpful abstractions to make manual instrumentation of agents simpler.

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.

If you'd prefer to use pure OTEL instead, see Setup using base OTEL

Installation

Ensure you have OpenInference and Phoenix OTEL installed:

pip install arize-phoenix-otel

npm install @arizeai/phoenix-otel @arizeai/openinference-core

For detailed API documentation, consult the respective documentation sites.

Arize Phoenix TSarize-ai.github.io

OpenInference JSarize-ai.github.io

Setting Up Tracing

from phoenix.otel import register

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

Using Helpers

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

1. Tracing a function

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

import { traceChain } from "@arizeai/openinference-core";

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

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

tracedFunc("input");

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

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/openinference-core";
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",
  }
);

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.

Chains

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-overriden-name")
def this_name_should_be_overriden(input: str) -> Dict[str, Any]:
    return {"output": "output"}

this_name_should_be_overriden("input")

Using Wrappers

import { withSpan } from "@arizeai/openinference-core";
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/openinference-core";

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/openinference-core";

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/openinference-core";

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

thisNameShouldBeOverriden("input");

Agents

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/openinference-core";
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/openinference-core";

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

decoratedAgent("input");

Tools

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-overriden-name",
    description="overriden-tool-description",
)
def this_tool_name_should_be_overriden(input1: str, input2: int) -> None:
    """
    this tool description should be overriden
    """

this_tool_name_should_be_overriden("input1", 1)

Using Function Wrappers

import { traceTool } from "@arizeai/openinference-core";

/**
 * 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/openinference-core";

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

thisToolNameShouldBeOverriden("input1", 1);

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.

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-4",
            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. The openinference-instrumentation library contains helper functions that produce valid OpenInference attributes for LLM spans:

get_llm_attributes
get_input_attributes
get_output_attributes

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-4",
    ),
    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-4",
    ),
    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-4",
)

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-4",
):
    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/openinference-core";
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/openinference-core 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/openinference-core";
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/openinference-core";
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-4", 0.5);

Additional Features

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

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 { suppressTracing } from "@opentelemetry/core";
import { withSpan } from "@arizeai/openinference-core";
import { trace, context } 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",
    }
  );
});

Using Context 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/openinference-core";
import { withSpan } from "@arizeai/openinference-core";
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",
      }
    );
  }
);

Adding Images to your Traces

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

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/openinference-core";
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",
  }
);

PreviousSetup using base OTEL NextSetup Projects

Last updated 3 days ago

Was this helpful?