OpenAI OpenTelemetry Integration

Overview

This integration provides support for using OpenTelemetry with the OpenAI framework. It enables tracing and monitoring of applications built with OpenAI.

Installation

1. Install traceAI OpenAI

pip install traceAI-openai

Set environment variables

Set the Future AGI keys for tracing and the OpenAI key for the Quickstart request:

export FI_API_KEY="your-future-agi-api-key"
export FI_SECRET_KEY="your-future-agi-secret-key"
export OPENAI_API_KEY="your-openai-api-key"

Ollama and self-hosted OpenAI-compatible clients

Keep using this package when your application calls Ollama through the OpenAI SDK. The official https://ollama.com/v1 endpoint is attributed to ollama automatically. A local address can host many providers, so set the provider explicitly for each local or self-hosted client:

from openai import OpenAI
from traceai_openai import OpenAIInstrumentor, set_client_provider

# Use the trace provider configured in the quickstart below.
OpenAIInstrumentor().instrument(tracer_provider=trace_provider)

client = OpenAI(base_url="http://localhost:11434/v1", api_key="ollama")
set_client_provider(client, provider="ollama")

The override controls gen_ai.provider.name on sync and async requests, including streams, embeddings, Responses and failures. It changes tracing metadata only; it adds no spans and sends no provider setting to the server. Use a lowercase provider identifier (1–64 characters, starting with a letter; letters, digits, ., _ and - are accepted), not a URL or secret.

Set the override before making requests. It takes precedence over host detection; set_client_provider(client, provider=None) restores host detection. Other clients are unaffected. A client returned by copy() or with_options() is a separate instance: set its override explicitly too, especially if its endpoint changes. Unconfigured loopback, proxy and unknown hosts retain the historical openai label; the instrumentor never assumes a local server is Ollama.

For applications using the native ollama Python client instead, use traceai-ollama. Do not install both instrumentors for the same API call. See Ollama's OpenAI compatibility guide for supported endpoints and models.

Quickstart

Register Tracer Provider

Set up the trace provider to establish the observability pipeline. The trace provider:

from fi_instrumentation import register
from fi_instrumentation.fi_types import ProjectType

trace_provider = register(
    project_type=ProjectType.OBSERVE,
    project_name="openai_app"
)

Configure OpenAI Instrumentation

Set up your OpenAI client with built-in observability. This includes support for text, image, and audio models.

from traceai_openai import OpenAIInstrumentor

OpenAIInstrumentor().instrument(tracer_provider=trace_provider)

Make a traced request

Create the client after calling OpenAIInstrumentor().instrument(...).

from openai import OpenAI

# OpenAI reads OPENAI_API_KEY from your environment.
with OpenAI() as client:
    response = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[
            {"role": "system", "content": "You are a helpful assistant."},
            {"role": "user", "content": "Can you tell me a joke?"},
        ],
    )
    print((response.choices[0].message.content or "").strip())
trace_provider.force_flush()  # Flush before a short-lived script exits.

Azure OpenAI

Use the same traceAI-openai package and OpenAIInstrumentor for Azure OpenAI. Keep the tracer registration and instrumentation above, and replace the OpenAI client and request with one of the following examples. Do not enable a second OpenAI instrumentor for the same calls.

Set your Azure resource key and deployment name (not the underlying model ID). You do not need an OpenAI API key:

export AZURE_OPENAI_API_KEY="your-azure-resource-key"
export AZURE_OPENAI_DEPLOYMENT="your-deployment-name"

v1 API

For a resource that supports the Azure OpenAI v1 API, set its complete base URL, including /openai/v1/:

export AZURE_OPENAI_BASE_URL="https://your-resource.openai.azure.com/openai/v1/"

import os
from openai import OpenAI

with OpenAI(
    base_url=os.environ["AZURE_OPENAI_BASE_URL"],
    api_key=os.environ["AZURE_OPENAI_API_KEY"],
) as client:
    response = client.chat.completions.create(
        model=os.environ["AZURE_OPENAI_DEPLOYMENT"],
        messages=[{"role": "user", "content": "Say hello in five words."}],
    )
    print(response.choices[0].message.content)
trace_provider.force_flush()

Versioned Azure API

For an existing AzureOpenAI application, keep your resource endpoint and an API version supported by your deployment. Public Azure resources use https://your-resource.openai.azure.com; Azure Government resources use https://your-resource.openai.azure.us:

export AZURE_OPENAI_ENDPOINT="https://your-resource.openai.azure.us"
export AZURE_OPENAI_API_VERSION="<supported-api-version>"

import os
from openai import AzureOpenAI

with AzureOpenAI(
    azure_endpoint=os.environ["AZURE_OPENAI_ENDPOINT"],
    api_version=os.environ["AZURE_OPENAI_API_VERSION"],
    api_key=os.environ["AZURE_OPENAI_API_KEY"],
) as client:
    response = client.chat.completions.create(
        model=os.environ["AZURE_OPENAI_DEPLOYMENT"],
        messages=[{"role": "user", "content": "Say hello in five words."}],
    )
    print(response.choices[0].message.content)
trace_provider.force_flush()

Both clients use the existing instrumentation, including async clients and streaming. Spans from resource hosts under openai.azure.com and openai.azure.us are labelled gen_ai.provider.name="azure"; custom gateways retain their existing provider label. If you use Microsoft Entra ID, keep the SDK token-provider configuration for your cloud instead of the API-key setup.

For AzureOpenAI and AsyncAzureOpenAI instances, including custom gateways, spans also record azure.openai.deployment (the client's azure_deployment, otherwise the request's model), azure.openai.api_version (when present in the client's default query), server.address (lowercase endpoint host only), and server.port (when explicitly present in the endpoint URL). Unavailable values are omitted; this metadata excludes full URLs, credentials and other query parameters. Plain OpenAI/AsyncOpenAI clients, including the v1 example above, keep provider attribution only. Explicit set_client_provider(...) overrides still take precedence for gen_ai.provider.name.

Azure Government does not change the trace export destination. Configure an approved Future AGI destination and content settings before tracing sensitive workloads; prompts and responses are captured by default.

Baseten

Use the same OpenAIInstrumentor with Baseten's OpenAI-compatible clients; no second instrumentor is needed. After the Quickstart tracer setup, create your client with api_key=os.environ["BASETEN_API_KEY"] and the appropriate base URL:

• Model APIs: https://inference.baseten.co/v1.
• Dedicated deployments: your deployment's OpenAI-compatible base URL, such as https://model-abc123.api.baseten.co/environments/production/sync/v1.

The hosted endpoint and dedicated model-*.api.baseten.co endpoints, including regional model endpoints, are attributed to baseten. Unknown proxies retain openai unless you set an explicit client provider override. Supported API routes and model names depend on the deployment; use the values in Baseten's inference API reference. The standalone Model APIs recipe in python/examples/baseten accepts only Model APIs URLs; use the OpenAI client directly for a dedicated deployment. Baseten's separate Performance Client and non-OpenAI protocols need their own tracing setup.

Vercel AI Gateway

Use the same OpenAIInstrumentor for Chat Completions sent to Vercel AI Gateway; no gateway-specific package or OpenAI key is needed. Set AI_GATEWAY_API_KEY to your Vercel AI Gateway key and AI_GATEWAY_MODEL to a current model ID:

export AI_GATEWAY_API_KEY="your-vercel-ai-gateway-key"
export AI_GATEWAY_MODEL="<provider>/<model>"

Keep the tracer registration and instrumentation from the Quickstart above, then replace its OpenAI client and request with:

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://ai-gateway.vercel.sh/v1",
    api_key=os.environ["AI_GATEWAY_API_KEY"],
)
response = client.chat.completions.create(
    model=os.environ["AI_GATEWAY_MODEL"],
    messages=[{"role": "user", "content": "Say hello in five words."}],
)
print(response.choices[0].message.content)
trace_provider.force_flush()  # Flush before a short-lived script exits.

The documented gateway host is labelled gen_ai.provider.name="vercel". This identifies the service the client calls, not the upstream model provider or a fallback provider. Unknown proxies and look-alike hosts retain their existing provider label. Sync, async, streamed and failed calls use the same attribution.

Client spans capture prompts, responses and returned token usage. They do not reconstruct gateway routing, retries, cost or provider attempts. Vercel's separate Trace Drains provide those gateway details but omit prompts and completions. A drain is not configured by OpenAIInstrumentor; do not sum client and drain token counts for the same request as separate model usage.