OpenAIException - error - 'StreamWrapper' object has no attribute 'parse'

[codefromthecrypt]

Describe the bug: ...

There's a crash when LiteLLM uses OpenAI (found in aider)

litellm.APIError: APIError: OpenAIException - error - 'StreamWrapper' object has no attribute 'parse', Received response - <opentelemetry.instrumentation.openai.wrappers.StreamWrapper object at 0x1776b36b0>, Type of response - <class 
'opentelemetry.instrumentation.openai.wrappers.StreamWrapper'>

To Reproduce

1. configure aider like this https://github.com/Aider-AI/aider/blob/main/CONTRIBUTING.md#handy-opinionated-setup-commands-for-macos--linux
2. configure edot like this https://github.com/elastic/elastic-otel-python/tree/main/examples/openai
3. update to latest openai instrumentation like this: pip install elastic-opentelemetry-instrumentation-openai==0.6.1
4. run dotenv run -- opentelemetry-instrument aider --model gpt-4o-mini --no-gitignore --no-show-model-warnings -m "how many fingers am I holding up?"

$ dotenv run -- opentelemetry-instrument aider --model gpt-4o-mini --no-gitignore --no-show-model-warnings -m "how many fingers am I holding up?"
───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
Aider v0.75.3.dev265+gbdefb1d5
Model: gpt-4o-mini with whole edit format
Git repo: .git with 451 files
Repo-map: using 4096 tokens, auto refresh

                    
litellm.APIError: APIError: OpenAIException - error - 'StreamWrapper' object has no attribute 'parse', Received response - <opentelemetry.instrumentation.openai.wrappers.StreamWrapper object at 
0x10de9dd30>, Type of response - <class 'opentelemetry.instrumentation.openai.wrappers.StreamWrapper'>

Environment (please complete the following information)

• OS: MacOS
• Python version: 3.12.8
• Framework and version [e.g. openai 1.63.0]: openai 1.63.0
• Instrumentation version: elastic-opentelemetry-instrumentation-openai==0.6.1

Additional context

I also tried ollama and same crash. It doesn't crash if you skip opentelemetry-instrument

[xrmx]

I think a fix for this would be switching StreamWrapper to be an instance of wrapt's ObjectProxy.

Said that if I'm looking at the right code I don't see any parse attribute there

[xrmx]

So the issue with missing tracing is reproducible with using the litellm library:

from litellm import completion

response = completion(
  model="openai/gpt-4o-mini",
  messages=[{ "content": "Hello, how are you?","role": "user"}],
  stream=True,
)

for event in response:
    print(event)

[codefromthecrypt]

Using latest main...

good news is that I don't get the error from aider anymore!
bad news is that I don't see openai spans, just HTTP ones

from a checkout of aider

Add .env

OPENAI_API_KEY=sk-REDACTED

OTEL_SERVICE_NAME=aider

# Change to 'false' to hide prompt and completion content
OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT=true
# Change to affect behavior of which resources are detected. Note: these
# choices are specific to the language, in this case Python.
OTEL_EXPERIMENTAL_RESOURCE_DETECTORS=process_runtime,os,otel,telemetry_distro

# Export metrics every 3 seconds instead of every minute
OTEL_METRIC_EXPORT_INTERVAL=3000
# Export traces every 3 seconds instead of every 5 seconds
OTEL_BSP_SCHEDULE_DELAY=3000

Setup pyenv

python3 -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install elastic-opentelemetry 'python-dotenv[cli]'
pip install -r requirements.txt
pip install .

bootstrap, replacing openai with main

edot-bootstrap --action=install
pip uninstall -y elastic-opentelemetry-instrumentation-openai
pip install 'elastic-opentelemetry-instrumentation-openai @ git+https://github.com/elastic/elastic-otel-python-instrumentations.git@main#subdirectory=instrumentation/elastic-opentelemetry-instrumentation-openai'

run aider

dotenv run -- opentelemetry-instrument aider --model gpt-4o-mini --no-gitignore --no-show-model-warnings -m "how many fingers am I holding up?"

[codefromthecrypt]

so I agree that if you run with the same dependencies of aider and use the python from #61 (comment) you get the same thing: http spans but no genai one