[Bug] - Document extraction fails with Ollama models for both PDF and TXT files for Synthetic Data Generation

[casualcomputer]

Describe the bug
I cannot extract documents using Ollama models. I've tried both PDF files and TXT files, and both fail to extract with the error "2 documents failed to extract. Retry or delete documents."

Checks

• I've read the troubleshooting guide (https://docs.kiln.tech/docs/troubleshooting-and-logs)
• I've tried to reproduce with another model (Qwen 3 VL Thinking 8B and Llama 3.2 11B Vision) - both fail
• I've searched the docs for a solution
• I've searched for existing Github issues/discussions - found related issue [Bug] - RAG Feature not working when using Ollama #814

To Reproduce
Steps to reproduce the behavior:

1. Go to 'Synthetic Data Generation'
2. Add PDF documents (doc1.pdf, doc2.pdf for try 1, doc1.txt, doc2.txt for try 2)
3. Click 'Retry Extraction'
4. Select extractor: 'Qwen 3 VL Thinking 8B (Vision-Language) (Ollama) - Markdown'
5. Click 'Run Extraction'
6. See error: "2 documents failed to extract"
7. Tried converting PDFs to TXT files
8. Changed extractor to 'Llama 3.2 11B (Vision) (Ollama) - Text'
9. Extraction still fails with same error

Expected behavior
Documents should extract successfully and show "Extracted" status, allowing me to proceed to Q&A generation.

Screenshots
The UI shows "2 documents failed to extract. Retry or delete documents." The TXT files show as "Extracted" in the list, but the error message persists.

System Information:

• OS: Win11
• Browser: Comet
• Kiln app Version: 0.23.0

Additional context
This appears to be related to issue #814 "[Bug] - RAG Feature not working when using Ollama". I've tried both PDF and TXT file formats with different Ollama vision models (Qwen 3 VL Thinking 8B and Llama 3.2 11B Vision), but extraction consistently fails.

[leonardmq]

@casualcomputer - Thanks for reporting the bug. This seems to be hardware dependent and we had a tentative fix / narrowing down of the issue.

Would you be able to try this dev build: https://github.com/Kiln-AI/Kiln/actions/runs/20927193826 (there are download links at the bottom of the page) and see if this is still happening?

Steps:

• close the normal / release app if you have it running
• install the dev build - since it is a dev build (unsigned), your OS may show some warnings

[casualcomputer]

I tried that dev build, but it didn't work. When I tried generating datasets, the app keeps spinning up new browser tabs (seemed to be stuck in an infinite loop, spinning up browser tabs). While it showed the document extraction process was ongoing, there was little GPU utilization; had the GPU been used, the GPU utilization would have been a lot higher.

[casualcomputer]

Update: I installed the Linux version on wsl on windows. I was able to create synthetic datasets on the Linux build inside windows wsl.