Research possibility of improving PDF handling in AI extensions using Baidu OCR model

[LukasHirt]

Right now we are quite limited with how we process PDF files in AI extensions. Using https://aiweekly.co/alerts/baidu-releases-mit-licensed-3b-ocr-model-for-long-documents might improve it.

[dj4oC]

Worth looking at Docling before committing to the Baidu model here.

The Baidu Unlimited-OCR model is genuinely interesting technically (MIT license, 3B params, single-pass multi-page parsing via Reference Sliding Window Attention), but it carries two practical problems for ownCloud deployments:
GPU dependency. It's a VLM-based model and needs a GPU for any reasonable throughput. Most oCIS self-hosted deployments (partners, public sector, SME) run on commodity VMs without NVIDIA hardware. That limits the realistic install base to a small minority of deployments.
Supply chain narrative. For European public sector customers especially, shipping a Baidu-trained model as part of the AI extensions pipeline is a conversation we'd rather not have. The MIT license covers legal reuse, but procurement reviews don't stop at license text.

Docling (https://github.com/docling-project/docling) addresses both:

CPU-native. Runs efficiently on commodity hardware without GPU. For digital-native PDFs (the majority of oCIS content) it's very fast. Scanned documents fall back to Tesseract internally.
MIT licensed. The toolkit itself is MIT; the Granite-Docling model (258M params) is Apache 2.0.
LF AI & Data Foundation project. Originated at IBM Research Zurich, now hosted under the Linux Foundation. Clean governance story for any procurement conversation.
Docker-native REST API. docling-serve ships pre-built container images and exposes a straightforward HTTP API. It drops into a Docker Compose stack alongside oCIS without any special infrastructure.
Format coverage. PDF, DOCX, PPTX, HTML, images. Structured Markdown output. Already integrated with LangChain and LlamaIndex, which is relevant if the AI extensions are doing RAG.

The output is clean Markdown with layout and table structure preserved, which is meaningfully better than what Tika/Tesseract pipelines typically produce for complex documents.
Rough integration sketch: a worker service subscribes to oCIS postprocessing events (NATS), catches uploads of PDF/Office files, POSTs to docling-serve, stores the extracted Markdown for search indexing or AI context injection. Docling-serve is stateless, so it scales horizontally without coordination.
Happy to sketch this out more concretely if useful. The European alternatives thread (LightOnOCR-2 from LightOn Paris, also Apache 2.0) is worth tracking too, but those require GPU and belong in a separate conversation.