ValueError: Circular reference detected with Gemini + file upload, and logs with large inline data blocking async event loop

[anthony2261]

Description

When google genai sdk is being logged, and a gemini call with a file upload is made, I get ValueError: Circular reference detected

Here's roughly how the code looks like:

# imports ... mainly what's needed
import io

import logfire
from google import genai
from google.genai import types
from pydantic import BaseModel

logfire.configure()
logfire.instrument_google_genai()

class TranscriptionResult(BaseModel):
    transcription: str
    reasoning: str

client = genai.Client(api_key=config.GOOGLE_GEMINI_API_KEY)

@logfire.instrument("Transcribe audio", extract_args=False)
async def transcribe_audio(audio_bytes: bytes)
    uploaded = client.files.upload(file=io.BytesIO(audio_bytes), config={"mime_type": "audio/wav"})
    gemini_response = await client.aio.models.generate_content(
        model="gemini-3-flash-preview",
        contents=["Transcribe this file", uploaded],
        config=types.GenerateContentConfig(
            response_mime_type="application/json",
            response_schema=TranscriptionResult,
            thinking_config=types.ThinkingConfig(thinking_budget=128),
        ),
    )
    return gemini_response.parsed

# Error happens when calling transcribe_audio

Full error log (with some files renamed):

ERROR:logfire:Caught an internal error in Logfire. Your code should still be running fine, just with less telemetry. This is just logging the internal error.
Traceback (most recent call last):
  File "/home/<project>/backend/.venv/lib/python3.11/site-packages/sentry_sdk/integrations/threading.py", line 123, in _run_old_run_func
    return old_run_func(self, *a[1:], **kw)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/<project>/backend/.venv/lib/python3.11/site-packages/anyio/_backends/_asyncio.py", line 967, in run
    result = context.run(func, *args)
             ^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/<project>/backend/src/services/analyze_audio.py", line 72, in _analyze_audio
    gemini_response = client.models.generate_content(
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/<project>/backend/.venv/lib/python3.11/site-packages/opentelemetry/instrumentation/google_genai/generate_content.py", line 1018, in instrumented_generate_content
    helper.process_request(contents, config, span)
  File "/home/<project>/backend/.venv/lib/python3.11/site-packages/opentelemetry/instrumentation/google_genai/generate_content.py", line 586, in process_request
    self._maybe_log_user_prompt(contents)
  File "/home/<project>/backend/.venv/lib/python3.11/site-packages/opentelemetry/instrumentation/google_genai/generate_content.py", line 830, in _maybe_log_user_prompt
    self._maybe_log_single_user_prompt(
  File "/home/<project>/backend/.venv/lib/python3.11/site-packages/opentelemetry/instrumentation/google_genai/generate_content.py", line 860, in _maybe_log_single_user_prompt
    self._otel_wrapper.log_user_prompt(
  File "/home/<project>/backend/.venv/lib/python3.11/site-packages/opentelemetry/instrumentation/google_genai/otel_wrapper.py", line 95, in log_user_prompt
    self._log_event(event_name, attributes, body)
  File "/home/<project>/backend/.venv/lib/python3.11/site-packages/opentelemetry/instrumentation/google_genai/otel_wrapper.py", line 110, in _log_event
    self._logger.emit(event)
  File "/home/<project>/backend/.venv/lib/python3.11/site-packages/logfire/_internal/integrations/google_genai.py", line 90, in emit
    span.add_event(record.event_name, attributes={'event_body': json.dumps(body, default=default_json)})
                                                                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.11/json/__init__.py", line 238, in dumps
    **kw).encode(obj)
          ^^^^^^^^^^^
  File "/usr/local/lib/python3.11/json/encoder.py", line 200, in encode
    chunks = self.iterencode(o, _one_shot=True)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.11/json/encoder.py", line 258, in iterencode
    return _iterencode(o, 0)
           ^^^^^^^^^^^^^^^^^
ValueError: Circular reference detected

Bonus: On a slow connection and for large audios, if I don't set extract_args=False, it blocks my async event loop (I'm using FastAPI) while it uploads the log. I even moved from gemini's inline audio upload to File API -> Gemini call because that also blocked the async loop while it was uploading. But not the main point of the bug.

Python, Logfire & OS Versions, related packages (not required)

logfire="4.32.1"
platform="Linux-6.6.87.2-microsoft-standard-WSL2-x86_64-with-glibc2.39"
python="3.11.11 (main, Mar 17 2025, 21:02:09) [Clang 20.1.0 ]"
[related_packages]
requests="2.33.1"
pydantic="2.11.7"
fastapi="0.136.1"
protobuf="6.33.6"
rich="14.1.0"
executing="2.2.0"
opentelemetry-api="1.40.0"
opentelemetry-exporter-otlp-proto-common="1.40.0"
opentelemetry-exporter-otlp-proto-http="1.40.0"
opentelemetry-instrumentation="0.61b0"
opentelemetry-instrumentation-asgi="0.61b0"
opentelemetry-instrumentation-fastapi="0.61b0"
opentelemetry-instrumentation-google-genai="0.7b0"
opentelemetry-instrumentation-httpx="0.61b0"
opentelemetry-instrumentation-requests="0.61b0"
opentelemetry-instrumentation-sqlalchemy="0.61b0"
opentelemetry-proto="1.40.0"
opentelemetry-sdk="1.40.0"
opentelemetry-semantic-conventions="0.61b0"
opentelemetry-util-genai="0.3b0"
opentelemetry-util-http="0.61b0"

[alexmojaki]

I'm not managing to reproduce. Can you try providing a complete self contained snippet that reproduces for you?

Bonus: On a slow connection and for large audios, if I don't set extract_args=False, it blocks my async event loop (I'm using FastAPI) while it uploads the log.

Logfire doesn't upload logs on the main thread. Extracting large args would cause some sync processing on the current thread and block the loop, but probably not for long, and it certainly wouldn't be affected by a slow connection. Whereas uploaded = client.files.upload(...) must happen synchronously since there's no await and absolutely depends on the network.

[anthony2261]

Thank you for your response, I'll prepare a proper snippet where I can replicate this exactly. Regarding client.files.upload, that was a copy/paste error, I meant to use client.aio.files.upload.

In any case, I'll ping you again once I have the snippet ready, thanks again.

[anthony2261]

Hey @alexmojaki , here's reproducible script:

logfire_repro.py

You'd need to run it with GOOGLE_GEMINI_API_KEY=... LOGFIRE_TOKEN=... AUDIO_PATH=./sample.wav uv run logfire_repro.py , then try to access http://localhost:8000/ping when Analyze Call prints in the logs. With a large enough audio file - like 10 minutes for example - you'll see that it stalls for ~10 seconds when tested on a slow internet connection (maybe even fast connections, not sure).

The stalling goes away if I set extract_args=False to the logfire.instrument decorator and if I comment out logfire.instrument_google_genai()

[alexmojaki]

I'm sure this is about processing the arguments, not uploading them. With a large enough file there will always be a problem. Trying to log that data in the span will always be a bad idea, it gets truncated anyway.

I was asking if you could repro the circular reference problem.

[anthony2261]

My bad. Will prepare something for the circular reference problem.

Regarding the arguments processing, is there a way not to log the inline audio bytes? Since I'm doing the following:

os.environ["OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT"] = "true"
logfire.instrument_google_genai()  # so, automatic logging of gemini calls with their contents
# ...
response = await client.aio.models.generate_content(
    # ...
    contents=[types.Part.from_bytes(data=audio_bytes, mime_type="audio/wav")],  # <-- automatically logged, problematic
)

Not sure if that's possible without having to go through the Gemini Files API

[alexmojaki]

I don't think there's currently a way with instrument_google_genai, a PR in logfire or the OTel instrumentation package would be needed.

You could maybe try https://pydantic.dev/docs/ai/core-concepts/direct/ instead of the google client, the pydantic-ai instrumentation supports specifically excluding binary content.