Reproduce false document labels when OCR returns no text
At a glance
A Paperless-ngx user documented how an AI classification workflow copied unrelated document metadata onto a blank image. This workflow is listed under Document Automation, Failure Analysis, AI Classification.
- Agents / tools
- Paperless-ngx 3.1.3, vLLM, Ollama
SourceBuilding a tool for these workflows? See weekly rates.
What they did
1) Use a test Paperless-ngx 3.1.3 installation with AI enabled, a chat-model backend and an embedding backend. The reporter used vLLM with an OpenAI-compatible endpoint and Ollama embeddings. 2) Import a normal text-bearing document so the index contains a retrievable neighbor. 3) Create a Document Added workflow with Apply AI suggestions, all fields selected, create missing objects enabled and overwrite existing values enabled. 4) Put a blank white JPEG into the consume folder and wait for OCR to finish with empty content. 5) Open the new document and compare its title, correspondent, type, date and tags with the existing document; inspect the workflow log alongside the OCR warning.
How it went
The reporter's blank JPEG inherited an unrelated PDF's metadata while the log reported successful application of AI suggestions.
Worth knowing
This is a historical failure reproduction, not an expected result across models. The closed thread links follow-up #13985.
