Bean_Sprout

**Update (May 2025):** I just pushed a fix that changes `DOCLING_SERVE_ENABLE_REMOTE_SERVICES` to default to `true`. Previously it defaulted to `false`, which blocked first-run model downloads for new installs. If you tried installing before and got model download errors, please try again with the updated template. If you prefer fully offline operation after your first successful start, you can toggle it back to `false`. [Camofox browser automation test — edit successful]

Same issue — the template was defaulting `DOCLING_SERVE_ENABLE_REMOTE_SERVICES` to `false`, which blocks model downloads. I just pushed a fix that changes the default to `true`. Please reinstall the template with the latest version and the models should download automatically on first boot. Since you switched from GPU to CPU, also make sure you select the `cpu` branch in the template (not just setting `DOCLING_DEVICE=cpu`). The `cpu` branch uses the `docling-serve-cpu` image which is built for CPU-only operation. Let me know how it goes!

This should be fixed now. I just pushed an update that changes `DOCLING_SERVE_ENABLE_REMOTE_SERVICES` to default to `true`, which allows first-run model downloads to work automatically. Please try reinstalling the template with the latest version. After your first successful start and the model cache is populated, you can set `DOCLING_SERVE_ENABLE_REMOTE_SERVICES` back to `false` if you prefer a fully offline setup. Let me know if it works!

I pushed a fix for our initial test. Let me know if the warning is resolved on your end.

I tracked this down and pushed a template update to address it. The warning: “artifacts_path is set to an invalid directory…” was caused by the model path mapping not matching Docling Serve’s expected artifacts path. What I changed: - Corrected the persistent models mount to: /opt/app-root/src/.cache/docling/models - Added explicit DOCLING_SERVE_ARTIFACTS_PATH support so it matches that path - Cleaned up template guidance around model storage/GPU behavior This warning can appear when the mounted appdata folder is empty/new (or mapped one level too high), because it can mask the image’s built-in model directory and trigger runtime downloads. After updating the template/container and applying settings, that warning should go away on normal runs. Thanks again for reporting it! y Your log snippet helped fix a real template issue.

Yo, forgot to quote you - see above!

Project: https://github.com/docling-project/docling-serve Template Repo: https://github.com/xxBeanSproutxx/unraid-docling-ca What is it? Docling is an open-source toolkit (IBM Research) that converts documents (PDF, DOCX, images, HTML, etc.) into structured Markdown or JSON. Perfect for AI workflows like RAG, all running locally on your Unraid server. Default Endpoints: - API: http://SERVER_IP:PORT - Docs: http://SERVER_IP:PORT/docs - Web UI: http://SERVER_IP:PORT/ui (set DOCLING_SERVE_ENABLE_UI=1) Install Notes: - First run may download models; caches are stored in appdata volumes. - CPU works out of the box. - For GPU, install the NVIDIA Driver plugin, reboot, choose a CUDA branch (cu128 or cu126) and add --gpus all. Volumes: - /mnt/user/appdata/docling/huggingface_cache → /opt/app-root/src/.cache/docling - /mnt/user/appdata/docling/easyocr_cache → /opt/app-root/.EasyOCR Issues? Post here if you run into setup questions or bugs.