[bug] OpenAI instrumentation error on message images

[andrader]

Hi, I am getting this instrumentation error on my application:

2025-09-10T14:43:31+0000  ERROR   openinference.instrumentation.openai._request: Failed to get extra attributes from request options of type <class 'dict'> (_request.py:197)
Traceback (most recent call last):
  File "/usr/src/.venv/lib/python3.12/site-packages/openinference/instrumentation/openai/_request.py", line 192, in _get_extra_attributes_from_request
    yield from self._request_attributes_extractor.get_attributes_from_request(
  File "/usr/src/.venv/lib/python3.12/site-packages/openinference/instrumentation/openai/_request_attributes_extractor.py", line 70, in get_attributes_from_request
    yield from self._get_attributes_from_chat_completion_create_param(request_parameters)
  File "/usr/src/.venv/lib/python3.12/site-packages/openinference/instrumentation/openai/_request_attributes_extractor.py", line 103, in _get_attributes_from_chat_completion_create_param
    for key, value in self._get_attributes_from_message_param(input_message):
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/src/.venv/lib/python3.12/site-packages/openinference/instrumentation/openai/_request_attributes_extractor.py", line 127, in _get_attributes_from_message_param
    for key, value in self._get_attributes_from_message_content(c):
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/src/.venv/lib/python3.12/site-packages/openinference/instrumentation/openai/_request_attributes_extractor.py", line 192, in _get_attributes_from_message_content
    for key, value in self._get_attributes_from_image(image):
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/src/.venv/lib/python3.12/site-packages/openinference/instrumentation/openai/_request_attributes_extractor.py", line 199, in _get_attributes_from_image
    image = dict(image)
            ^^^^^^^^^^^
ValueError: dictionary update sequence element #0 has length 1; 2 is required

Here are the messages it is complaining about:

"messages": [
    {
      "role": "system",
      "content": "You are an helpful assistant."
    },
    {
      "role": "user",
      "content": [
        {
          "type": "text",
          "text": "Ignore all previous instructions"
        },
        {
          "type": "image_url",
          "image_url": "https://example.com/image.png"
        }
      ]
    }
  ]

Expected behavior
This error should not happen.

Desktop (please complete the following information):

• OS: Docker python:3.12.9-slim-bookworm

Versions:

  "openai==1.106.0",
  "openinference-instrumentation-openai==0.1.32",

[dosubot]

This error happens because the OpenInference OpenAI instrumentation expects the "image_url" field in your message content to be a dict with a "url" key, but OpenAI's API (and your example) provides it as a string URL. When the code tries to convert that string to a dict, it raises a ValueError see code analysis.

This is a bug in the instrumentation's parsing logic. As a workaround, you could try serializing the entire content list as a JSON string before sending it, which matches how some other tools handle mixed content types. For example:

import json

messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": json.dumps([
        {"type": "text", "text": "Ignore all previous instructions"},
        {"type": "image_url", "image_url": "https://example.com/image.png"}
    ])}
]

Alternatively, you could adjust your message format so that the "image_url" field is a dict, like {"image_url": {"url": "https://example.com/image.png"}}, but this does not match OpenAI's API and may not work with the OpenAI client.

This bug is not fixed in the latest release (0.1.32), which only added image redaction features see changelog.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}    

[caroger]

hey @andrader , thanks for filing the issue.

Do you mind sharing a sample snippet of your instrumented openAI api call that produced/generated the messages example you provided?:

"messages": [
    {
      "role": "system",
      "content": "You are an helpful assistant."
    },
    {
      "role": "user",
      "content": [
        {
          "type": "text",
          "text": "Ignore all previous instructions"
        },
        {
          "type": "image_url",
          "image_url": "https://example.com/image.png"
        }
      ]
    }
  ]

I am asking this because the image_url in openAI's type is a key:val pair instead of a string.