google-gemini
`response.text` is no longer directly accessible - `TypeError: argument of type 'Part' is not iterable`
Description of the bug:
Accessing the property response.text no longer works.
genai.configure(api_key = gcp_key)
model = genai.GenerativeModel("gemini-pro")
response = model.generate_content(prompt)
print(str(response.text)
# TypeError: argument of type 'Part' is not iterable
# google/generativeai/types/generation_types.py
# line 327, in text
# if len(parts) != 1 or "text" not in parts[0]:
As a workaround, I can get the detail with response.candidates[0].content.parts[0].text, but this seems like it might be a breaking change?
Actual vs expected behavior:
I should be able to reference response.text directly.
Any other information you'd like to share?
NOTE: I am using the internal Google3 version of this SDK - see go/generative-ai-python-sdk-response-type-issue for the exact line internally. There seems to be a discrepancy between Google3 and GitHub - this is the line in the GitHub repo.
Feel free to reach out to me internally - mitchspano@
The public python examples are okay, this is a google-internal problem.
@mitchspano I've added you to an email thread titled "Checking for oneof field presence in UPB proto messages", where we're discussing the cause of this.
Thanks @MarkDaoust .. I am getting below error using langchain aswell as google-cloud-aiplatform. Am i missing something.
Interesting. Thanks for the update. Reopened.
Internally it has something to do with switching to upb protos, IDK if it's the same problem here.
https://github.com/protocolbuffers/upb?tab=readme-ov-file https://github.com/protocolbuffers/protobuf/tree/main/upb
proto-plus (used by these libraries)depends on an older version of protobuf (<5, when latest is 5.26.1), but upgrading to 5.26 doesn't cause this error.
Can ou post a "pip freeze | grep proto"?
Just with the python versions, using this SDK I can't reproduce that.
Can you reproduce the error in Colab, for example, so we can replicate it and share with the owners?
Used below to create my environment where i see the issue.
name: python-test channels:
- defaults dependencies:
- python
- pip=23.3
- requests=2.31
- pandas=2.2
- plotly=5.19
- sqlalchemy=2.0
- streamlit=1.32
- fastapi=0.103
- uvicorn=0.20
- ipykernel=6.28
- pip:
- duckdb == 0.10
- google-cloud-aiplatform
- langchain
- langchain-google-vertexai
The issue happens when parts is a protobuf object and not a dict - the check needs to be replaced with:
if len(parts) != 1 or not hasattr(parts[0], "text"):
Instead of:
if len(parts) != 1 or "text" not in parts[0]:
I am not sure if it will resolve all issues or if it needs to be a more comprehensive check, but it was causing issues for me and I ended up using parts[0].text directly to avoid similar issues.
https://github.com/google-gemini/generative-ai-python/blob/419a7cef81188a2d715b3853113a3213516d6a4a/google/generativeai/types/generation_types.py#L416
For some reason - I was getting this error consistently yesterday, but can't reproduce today!!!
@sakher, you do have a point, hasattr(parts[0], "text") would be more robust here. But the problem is that "text" in parts[0] is supposed to work for the proto-plus based proto classes. This library is only setup to work with those. Switching to hasattr will just mask this error and move the problem somewhere else. There are a lot of isinstance checks that will go the wrong way.
You guys know better, I haven't dug deep into it - but when the error happens (sporadically) from some debugging, the type of parts was:
google.protobuf.pyext._message.RepeatedCompositeContainer
However, in some cases, the error doesn't happen, and the type of parts seems to be different in such cases:
proto.marshal.collections.repeated.RepeatedComposite
(@MarkDaoust see me inside corp as parrt)
I'm also seeing this with anaconda and mac os (Sonoma 14.5) on apple silicon via jupyter lab.
TypeError: argument of type 'Part' is not iterable
Snippet:
import google.generativeai as genai
genai.configure(api_key='SECRET API KEY')
code = """
int i = 34;
"""
model = genai.GenerativeModel(model_name="gemini-1.5-flash")
response = model.generate_content([
"""You are an expert Java programmer.
Here is some Java code that I would like you to analyze.""", code])
print(response.text)
Oddly however, the response object does appear to have the appropriate elements:
GenerateContentResponse(
done=True,
iterator=None,
result=protos.GenerateContentResponse({
"candidates": [
{
"content": {
"parts": [
{
"text": "This code snippet is a simple declaration
...
)
For example the following evaluates fine: response.candidates[0].content.parts[0].text:
"This code snippet is a simple declaration ...
Info:
$ conda list python
# packages in environment at /Users/parrt/opt/anaconda3:
#
# Name Version Build Channel
antlr4-python3-runtime 4.12.0 pypi_0 pypi
brotli-python 1.0.9 py39hcec6c5f_8
google-api-python-client 2.133.0 pypi_0 pypi
ipython 8.15.0 py39hecd8cb5_0
ipython_genutils 0.2.0 pyhd3eb1b0_1
msgpack-python 1.0.3 py39haf03e11_0
python 3.9.19 h5ee71fb_1
python-dateutil 2.9.0post0 py39hecd8cb5_2
python-fastjsonschema 2.16.2 py39hecd8cb5_0
python-graphviz 0.20.1 pypi_0 pypi
python-libarchive-c 2.9 pyhd3eb1b0_1
python-lmdb 1.4.1 py39hcec6c5f_0
python-lsp-black 2.0.0 py39hecd8cb5_0
python-lsp-jsonrpc 1.1.2 pyhd3eb1b0_0
python-lsp-server 1.10.0 py39hecd8cb5_0
python-slugify 5.0.2 pyhd3eb1b0_0
python-snappy 0.6.1 py39hcec6c5f_0
python.app 3 py39hca72f7f_0
python_abi 3.9 2_cp39 conda-forge
$ conda list anaconda
# packages in environment at /Users/parrt/opt/anaconda3:
#
# Name Version Build Channel
_anaconda_depends 2022.05 py39_0
anaconda custom py39_2
anaconda-anon-usage 0.4.4 py39hfd8fc91_0
anaconda-client 1.12.3 py39hecd8cb5_0
anaconda-navigator 2.1.4 py39hecd8cb5_0
anaconda-project 0.11.1 py39hecd8cb5_0
My environment was old so I whacked anaconda and reinstalled everything. I get the same problem still.
$ conda list python
# packages in environment at /Users/parrt/anaconda3:
#
# Name Version Build Channel
brotli-python 1.0.9 py312h313beb8_8
gitpython 3.1.37 py312hca03da5_0
google-api-python-client 2.135.0 pypi_0 pypi
ipython 8.25.0 py312hca03da5_0
ipython_genutils 0.2.0 pyhd3eb1b0_1
msgpack-python 1.0.3 py312h48ca7d4_0
python 3.12.4 h99e199e_1
python-dateutil 2.9.0post0 py312hca03da5_2
python-dotenv 0.21.0 py312hca03da5_0
python-fastjsonschema 2.16.2 py312hca03da5_0
python-json-logger 2.0.7 py312hca03da5_0
python-libarchive-c 2.9 pyhd3eb1b0_1
python-lmdb 1.4.1 py312h313beb8_0
python-lsp-black 2.0.0 py312hca03da5_0
python-lsp-jsonrpc 1.1.2 pyhd3eb1b0_0
python-lsp-server 1.10.0 py312hca03da5_0
python-slugify 5.0.2 pyhd3eb1b0_0
python-snappy 0.6.1 py312h313beb8_0
python-tzdata 2023.3 pyhd3eb1b0_0
python.app 3 py312h80987f9_0
$ conda list anaconda
# packages in environment at /Users/parrt/anaconda3:
#
# Name Version Build Channel
_anaconda_depends 2024.06 py312_openblas_2
anaconda-anon-usage 0.4.4 py312hd6b623d_100
anaconda-catalogs 0.2.0 py312hca03da5_1
anaconda-client 1.12.3 py312hca03da5_0
anaconda-cloud-auth 0.5.1 py312hca03da5_0
anaconda-navigator 2.6.0 py312hca03da5_0
anaconda-project 0.11.1 py312hca03da5_0
I think this is the fix we need: https://github.com/googleapis/proto-plus-python/pull/471