Hi Unraid community!

I'm planning to build a family document management system on my Unraid server and would love to get your input on the architecture and OCR implementation before I dive in.

## Hardware & Current Setup

Server:

- Aoostar WTR PRO MAX

- AMD Ryzen 9 7940HS (8C/16T)

- AMD Radeon 780M iGPU

- 32GB RAM DDR5 (expandable to 64GB if needed)

- 3x Samsung 990 PRO 1TB NVMe (Cache RAID1 + SSD pool)

- 2x 8TB HDD (Parity + Data)

- Unraid 7.2.3

Currently running:

- Immich (photo management with AI/ML using the iGPU)

- Jellyfin (media streaming with VAAPI hardware transcoding)

- Tailscale (remote access)

## Planned System Goals

Use case:

- Family of 3 (me, wife, daughter)

- ~10-20 documents per month (invoices, contracts, medical records, receipts)

- Mix of document types:

- 50% electronic PDFs (invoices from utilities, banks)

- 30% phone camera scans (receipts, paper documents)

- 20% medical documents (some handwritten notes)

What I want to achieve:

1. Easy upload from phones (iOS/Android)

2. OCR with Polish + English language support

3. Smart tagging/classification (invoices, medical, contracts, etc.)

4. Table extraction from invoices (for potential budget automation)

5. Full-text search through archived documents

6. Long-term archival with metadata

## Proposed Architecture

I'm considering a hybrid setup:

NEXTCLOUD

- File sync & temporary upload hub

- Mobile apps for easy document upload

- Office file preview/editing (optional Collabora)

- Folder: /Scan/Inbox → auto-consumed by Paperless

PAPERLESS-NGX

- Document management system (DMS)

- OCR & archival

- AI-powered tagging & classification

- Consume folder monitors Nextcloud uploads

- Full-text search & metadata management

Database layer:

- Separate PostgreSQL + Redis for each stack (isolated Docker networks)

- Similar to my existing Immich setup

## The OCR Question - This is where I need your advice!

I'm torn between different OCR approaches:

### Option 1: Built-in Tesseract OCR (simple)

Pros:

- Zero additional setup

- Works out of the box with Paperless

- Good for clean PDFs

- Low RAM usage (~50-100MB)

Cons:

- Struggles with table structure (loses layout)

- Poor handwriting recognition

- Medium quality on phone camera scans

### Option 2: DeepSeek OCR v2 / AI-powered OCR (advanced)

Pros:

- Excellent table structure preservation

- Good handwriting recognition (medical notes!)

- Better accuracy on poor quality scans

- Context-aware processing

Cons:

- Requires custom Docker implementation

- AMD GPU (ROCm) support may be tricky

- More RAM (~2-3GB)

- Maintenance overhead on Paperless updates

- No official Docker image available

### Option 3: Alternative AI OCR (middle ground?)

Someone mentioned Surya OCR or PaddleOCR as potentially easier alternatives with Docker images and AMD GPU support.

## Questions for the Community

1. Has anyone implemented Paperless-ngx with custom OCR backends?

- How did you integrate it?

- Was it worth the extra complexity?

2. AMD Radeon 780M iGPU users:

- Any success running ROCm-based OCR in Docker containers?

- Should I just stick to CPU inference for OCR?

3. Tesseract vs AI OCR real-world experience:

- For mixed document types (PDF + scans), is Tesseract's accuracy sufficient?

- Does table structure extraction really matter in practice?

4. Docker images recommendations:

- Any ready-to-use OCR service containers that work well with Paperless?

- Surya OCR, PaddleOCR, or other alternatives you'd recommend?

5. Architecture feedback:

- Does the Nextcloud → Paperless workflow make sense?

- Any gotchas I should be aware of?

- Alternative approaches I should consider?

## My Current Thinking

I'm leaning towards:

- Phase 1: Start with Tesseract (built-in) and test with real documents for 2-4 weeks

- Phase 2: Upgrade to AI OCR only if Tesseract proves insufficient

But if there's a ready-to-use Docker solution that works well with AMD GPUs and doesn't require heavy customization, I'd rather set it up correctly from the start.

## RAM Considerations

With 32GB RAM, current usage is:

- Unraid OS: ~2GB

- Immich stack: ~800MB

- Jellyfin: ~400MB

- Planned Nextcloud stack: ~600MB

- Planned Paperless stack: ~500MB

- Total: ~4.3GB (plenty of headroom)

Adding AI OCR would bring it to ~6-7GB total, which is still comfortable. I can expand to 64GB if needed.

---

TL;DR: Building Nextcloud + Paperless-ngx DMS on Unraid with AMD Radeon 780M iGPU. Seeking advice on OCR approach (Tesseract vs AI OCR), Docker implementations, and overall architecture feedback.

Thanks in advance for any insights! This community has been incredibly helpful so far.

Is there somebody that can help me with this?

Have you found a solution for your need? I am also in the same situation and trying to find solution for automatic tagging with the help of AI?

Follow Ed's recent Paperless series and customize it for your specific needs. https://www.youtube.com/watch?v=KeB-74pPqh8&list=PLv4nZVjyIOj1cyodwaHSVfvQWBC3ZtWDG