OpenInference OpenAI Instrumentation
Python auto-instrumentation library for OpenAI's python SDK.
The traces emitted by this instrumentation are fully OpenTelemetry compatible and can be sent to an OpenTelemetry collector for viewing, such as Arize Phoenix or Arize AX.
Installation
pip install openinference-instrumentation-openai
SDK compatibility
The supported SDK range is openai>=2.8.0, matching the instruments extra
and the OpenTelemetry instrumentation dependency check. OpenAI 1.x and earlier
2.x releases are outside this range. Install the SDK with the instrumentor using:
pip install 'openinference-instrumentation-openai[instruments]'
CI tests OpenAI 2.8.0, OpenAI 3.0.0, and the latest available SDK on Python 3.10 and 3.14. There is no SDK upper bound; the latest lane is a required compatibility check, including when a new major version becomes available.
OpenAI 3 uses httpx2 as its default HTTP transport. OpenAIInstrumentor wraps
the SDK request methods and supports this transport without application changes.
Chat Completions, Completions, Embeddings, Responses, Images, and Decisions are
tested with sync and async clients, including streaming where the API supports it.
OpenInference spans are independent of HTTP transport spans:
opentelemetry-instrumentation-httpx does not instrument the SDK's default httpx2
client in normal application use.
Decisions API
Calls to the Decisions API
(client.decisions.create, openai>=3.26.0) are recorded as DECISION spans
following the decision span conventions.
A decision model answers a fixed set of typed questions (predicate, choice,
score) about the input rather than generating text, so the span identifies the
model under decision.* instead of llm.* and carries no llm.* attributes:
| Attribute | Value |
|---|---|
openinference.span.kind |
DECISION |
decision.system |
openai |
decision.provider |
Inferred from the client host, e.g. openai or azure |
decision.request.model_name |
The model sent in the request |
decision.response.model_name |
The model reported in the response |
decision.model_name |
The response model, falling back to the requested model |
decision.token_count.input |
usage.input_tokens |
decision.token_count.output |
usage.output_tokens |
input.value / input.mime_type |
The request body (input, questions, model) as JSON |
output.value / output.mime_type |
The response body (answers, model, usage) as JSON |
The request and the answers live only in input.value and output.value, so the
hide_inputs and hide_outputs trace configuration flags redact them completely, and
hide_input_images / base64_image_max_length apply to inline images in the input.
See examples/decisions.py for a runnable example.
Development and testing
From the repository root, run the three SDK lanes (formatting, lint, types, and tests):
uvx --with tox-uv tox -c python/tox.ini run -e py310-ci-openai,py310-ci-openai-v3,py310-ci-openai-latest
Use py314 in place of py310 to run the other CI Python version. The baseline
SDK is pinned in test-requirements.txt; the v3 and latest overrides live in
python/tox.ini and follow the repository's dependency release-age policy.
The pytest-only _httpx2_compat plugin loads before HTTP mocking plugins. When
httpx2 is installed, it aliases httpx and httpcore to their version 2 modules,
so RESPX and VCR intercept the SDK's native transport. This does not substitute a
legacy HTTP client into the SDK. The alias also lets the test suite exercise HTTP
child spans; application instrumentation does not load this plugin.
Tests block network access and replay existing VCR cassettes by default. Missing
mocks or cassette entries fail instead of contacting OpenAI or Azure. To deliberately
record a cassette, run the selected test with --record-mode=once and valid provider
credentials. This enables network access for VCR-marked tests. Request and response
headers are stripped; review recorded bodies before committing them.
Quickstart
In this example we will instrument a small program that uses OpenAI and observe the traces via arize-phoenix.
Install packages.
pip install openinference-instrumentation-openai "openai>=2.8.0" arize-phoenix opentelemetry-sdk opentelemetry-exporter-otlp
Start the phoenix server so that it is ready to collect traces. The Phoenix server runs entirely on your machine and does not send data over the internet.
python -m phoenix.server.main serve
In a python file, setup the OpenAIInstrumentor and configure the tracer to send traces to Phoenix.
import openai
from openinference.instrumentation.openai import OpenAIInstrumentor
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk import trace as trace_sdk
from opentelemetry.sdk.trace.export import ConsoleSpanExporter, SimpleSpanProcessor
endpoint = "http://127.0.0.1:6006/v1/traces"
tracer_provider = trace_sdk.TracerProvider()
tracer_provider.add_span_processor(SimpleSpanProcessor(OTLPSpanExporter(endpoint)))
# Optionally, you can also print the spans to the console.
tracer_provider.add_span_processor(SimpleSpanProcessor(ConsoleSpanExporter()))
OpenAIInstrumentor().instrument(tracer_provider=tracer_provider)
if __name__ == "__main__":
client = openai.OpenAI()
response = client.chat.completions.create(
model="gpt-3.5-turbo",
messages=[{"role": "user", "content": "Write a haiku."}],
max_tokens=20,
stream=True,
stream_options={"include_usage": True},
)
for chunk in response:
if chunk.choices and (content := chunk.choices[0].delta.content):
print(content, end="")
Since we are using OpenAI, we must set the OPENAI_API_KEY environment variable to authenticate with the OpenAI API.
export OPENAI_API_KEY=your-api-key
Now simply run the python file and observe the traces in Phoenix.
python your_file.py
FAQ
Q: How to get token counts when streaming?
A: To get token counts when streaming, install openai>=2.8.0 and set stream_options={"include_usage": True} when calling create. See the example shown above. For more info, see here.
More Info
1 maintainer
Metadata
1 maintainer
Release files for openinference-instrumentation-openai 0.1.66
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | |
|---|---|---|---|---|
| openinference_instrumentation_openai-0.1.66.tar.gz | 30.5 kB | With provenance | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| openinference_instrumentation_openai-0.1.66-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 68.9 kB
Release files / openinference_instrumentation_openai-0.1.66.tar.gz
| Download URL | openinference_instrumentation_openai-0.1.66.tar.gz |
|---|---|
| Size | 30.5 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
2c39ffd053c4105fe9b96ad0f8a3884cbd949ee5b1a3ad9b20798fb9733f3bb2
|
|
BLAKE2b-256 checksum How to use checksums |
22175563b169a73730c5a0261bee39146b8f871e4719eb9a2b963162252bf79c
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 23, 2026.
Transparency logRelease files / openinference_instrumentation_openai-0.1.66-py3-none-any.whl
| Download URL | openinference_instrumentation_openai-0.1.66-py3-none-any.whl |
|---|---|
| Size | 38.4 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
cef7ea684f23eb3501f850fc8baebd84c034e17737d6e50172f1e87e2e4a1843
|
|
BLAKE2b-256 checksum How to use checksums |
ea7154de247e467f1ead9d2647e707c28db8cb485a87c14067b50dcc5a76c487
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 23, 2026.
Transparency logSecurity for openinference-instrumentation-openai 0.1.66
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
Full provenance
All 2 files in this release have been uploaded with provenance metadata.
Attestation for openinference_instrumentation_openai-0.1.66.tar.gz
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 23, 2026.
Transparency logA note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.
