[Bug]: Cannot parse images

[weblerson]

Self Checks

• I have searched for existing issues search for existing issues, including closed ones.
• I confirm that I am using English to submit this report (Language Policy).
• Non-english title submitions will be closed directly ( 非英文标题的提交将会被直接关闭 ) (Language Policy).
• Please do not modify this template :) and fill in all the required fields.

RAGFlow workspace code commit ID

de24e74

RAGFlow image version

de24e74(v0.21.1)

Other environment information

Actual behavior

ALWAYS when I try to parse a image itself or a PDF file with a image within it I get this error:

03:09:34 Task has been received.
03:09:35 Page(1100000001): Finish OCR: ( ...)
03:09:35 Page(1100000001): Use CV LLM to describe the picture.
03:09:36 Page(1100000001): [ERROR]cannot identify image file <_io.BytesIO object at 0x7d7b84415e90>
03:09:36 Page(1100000001): No chunk built from image.jpg

Anyone has faces the same issue as mine? Is there something I might be missing?

Expected behavior

No response

Steps to reproduce

- Create a Gemini provider with your API key
- Attach gemini-2.5-flash as image2text model
- Create a dataset
- Populate the dataset with image files
- Try to parse the images

Additional information

No response

[dosubot]

Hi @weblerson! I'm Dosu and I’m helping the ragflow team.

This error is usually caused by the image data being corrupted, in an unsupported format, or not properly loaded when RAGFlow tries to process it with PIL/Pillow. RAGFlow officially supports parsing images in jpg, jpeg, png, and gif formats for the image2text pipeline. The error can also occur if images extracted from PDFs are malformed or if the provider/model integration doesn't handle the image format as expected.

Recent versions require images for IMAGE2TEXT tasks to be passed as base64-encoded data URIs, and not all providers (including some Gemini setups) may support this format out of the box. If your provider doesn't support this, image parsing will fail. Also, there are known bugs in v0.21.1 that can interrupt PDF/image processing, such as logging errors and issues with image chunking, which may require patching the code to resolve temporarily (details, details).

If you are consistently seeing this error with supported image formats, double-check that the images are not corrupted and are being uploaded in one of the supported formats. If you are parsing PDFs, the embedded images must also be in a supported format and properly extracted. If you are using a custom provider or pipeline, ensure that images are converted to base64 data URIs before being sent to the model. If possible, try switching to a provider known to work with RAGFlow's image requirements (like Ollama) to see if the issue persists.

Recent merged fixes and enhancements to image parsing may address some of these issues in newer versions (details, details). If you are able to update, consider testing with the latest code.

To reply, just mention @dosu.

---

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

[weblerson]

All the images i've ingested on the RagFlow instance are valid images. I will try another image2text model and see if the error persists. Any other advice? Maybe a model that currently supports the base64-encoded out-of-the-box?

@dosu

[dosubot]

For models that support base64-encoded images out-of-the-box in RAGFlow v0.21.1, you can use Ollama, Gemini, Nvidia, Anthropic, and most OpenAI-compatible providers (like GptV4, AzureGptV4, xAICV, QWenCV, HunyuanCV, Zhipu4V, StepFunCV, VolcEngineCV, LmStudioCV, OpenAI_APICV, TogetherAICV, YiCV, SILICONFLOWCV, OpenRouterCV, LocalAICV, XinferenceCV, GPUStackCV). These models/classes handle the base64 conversion internally, so you don’t need to do anything special with your images—just make sure the provider is set up correctly and the images are valid and supported formats (jpg, png, etc.)
Ollama and Gemini are both good choices for testing this workflow, as they have explicit logic for base64 image handling in the codebase. If you still see errors, double-check for any local logging bugs or chunking issues in your version, as these can also interrupt image parsing workflows.
If you want to try another provider, switching to Ollama or Nvidia is a reliable way to confirm if the issue is provider-specific or a broader pipeline/config problem.
You can find the relevant implementation details in the codebase for reference: cv_model.py.

To reply, just mention @dosu.

---

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

[weblerson]

I am using Gemini. Still does not work. The same [ERROR]cannot identify image file <_io.BytesIO object at 0x724044da8130> error again and again.

@dosu

[dosubot]

This error is triggered when PIL (Pillow) tries to open image data from memory and can't recognize it as a valid image. In the GeminiCV pipeline, images are first converted to base64, then decoded and loaded with PIL. If the base64 data is malformed, truncated, or not actually an image (sometimes happens with PDF extraction), you'll get this exact error. Even if your original files are valid, something in the extraction or conversion step may be corrupting the data or passing non-image content downstream.

Recent updates to the GeminiCV model have improved image handling and fixed some parameter issues, but the core image loading logic still relies on PIL being able to recognize the decoded bytes as a valid image format (jpg, png, etc.)—if not, it fails at this step (source).

To troubleshoot:

• Try parsing a simple, small, standalone jpg or png (not from a PDF) to see if the error persists.
• If the error only happens with images extracted from PDFs, the extraction step may be producing unsupported or corrupted image data.
• Enable debug logging to capture the actual base64 string or image bytes being passed to PIL, and try opening them manually with PIL.Image.open to confirm validity.
• If you need Gemini specifically, consider updating to the latest codebase, as recent PRs have improved error handling and compatibility for GeminiCV (PR #10799).
• As a workaround, you can use the OpenAI-compatible provider with Gemini models, which avoids some Gemini-specific API issues (source).

If you still see the error with simple images, it's likely a pipeline bug or a config issue with how images are passed to GeminiCV. If it only happens with PDFs, the extraction logic may need review.

To reply, just mention @dosu.

---

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