Docling Studio – open-source visual inspection tool for Docling pipelines

[pjmalandrino]

Hi team and community 👋

I've been building on top of Docling for a production RAG system and hit a
recurring pain point: validating conversion outputs means digging through JSON
and mentally mapping bounding box coordinates back to the original pages.
No visual feedback loop.

So I built Docling Studio — an open-source visual inspection and debugging
layer for Docling.

🔗 GitHub: https://github.com/scub-france/Docling-Studio

What it does

• Upload a PDF, configure your pipeline (OCR engine, table extraction,
  enrichment options)
• Run the conversion
• Visually inspect every detected bounding box overlaid on the original page
• Explore interactive chunking output for RAG pipelines
• Toggle features dynamically via feature flags

Stack

• Frontend: Vue 3 / TypeScript
• Backend: FastAPI / Python (hexagonal ports & adapters architecture)
• Storage: SQLite
• Packaging: single multi-arch Docker image
• Conversion modes: local (Docling embedded) and remote (Docling Serve)

Why it might be useful

Docling delivers excellent extraction quality, but when a table or section is chunked incorrectly, understanding why can quickly become frustrating without proper visual tools.

I ran into this issue myself on several projects, which led me to build a tool to solve it.

Happy to get feedback, and open to contributions!
The roadmap includes a complete ingestion pipeline for RAG and a Hugging Face Space deployment.

[klcd]

Hey @pjmalandrino

Very cool project, love the idea and the architecture.

I've been playing around with for a bit and implemented the following ideas locally.

• Display the Element hierarchically so they can be folded and merge BBoxes of folded elements.
• Displaying the document graph directly in the verify section and scoping it to the active page.
• Toggling BBoxes and other small things

I am currently mostly thinking about how to implement a simple Editor for the Docling Document so the view can be used to correct the detected types, levels etc.

If this interests you as well I would be very happy to talk to you

[pjmalandrino]

Hey @klcd,

Thank you!
Sure I am interested, these are kind of sitting in the roadmap somewhere, so definitely worth integrating.

I am actually working on the "ingest" part to prepare the 1.x, so any ideas to improve the visual interactive part are very welcome.

You can contact me on LinkedIn (https://www.linkedin.com/in/pier-jean-malandrino/) or by mail as you prefer: pierjean.malandrino@scub.net