AI-Powered Document Parser | PDF to JSON | OCR + LLM Pipeline
Extract structured data from PDFs and images using DeepSeek-OCR-2 and local LLMs.
100% local processing. No cloud APIs. Your data stays private.
VisionIngest is a local-first document parsing application that converts PDFs, scanned documents, and images into structured JSON data. It uses a two-stage AI pipeline:
- Vision Model (Eyes): DeepSeek-OCR-2 extracts text and layout as markdown
- Language Model (Brain): Local LLM (via Ollama) converts markdown to structured JSON
No data leaves your machine. Perfect for sensitive documents like resumes, contracts, invoices, and medical records.
Upload interface with document type selection, processing modes, and OCR quality control
Side-by-side view: Original document vs. Extracted JSON (editable)
| Feature | Description |
|---|---|
| Multi-Document Support | Resumes, invoices, receipts, contracts, business cards, and more |
| 100% Local Processing | No cloud APIs - your data never leaves your machine |
| GPU Accelerated | CUDA support for fast processing on NVIDIA GPUs |
| Flexible Processing Modes | Performance, Batch, and Low VRAM modes for any hardware |
| Adjustable OCR Quality | 4 presets from fast (512px) to high-quality (1280px) |
| Editable JSON Output | Edit extracted data directly in the UI |
| REST API | Full API for integration with other applications |
| Modern Web UI | React-based interface with real-time GPU monitoring |
- HR & Recruiting: Resumes, CVs, Cover Letters
- Finance: Invoices, Receipts, Bank Statements
- Legal: Contracts, NDAs, Agreements
- Business: Business Cards, Reports, Forms
- Academic: Transcripts, Certificates, Research Papers
- Custom: Add your own templates with JSON schema + prompt
| Component | Technology |
|---|---|
| Vision Model | DeepSeek-OCR-2 (Visual Causal Flow architecture) |
| Language Model | Local LLM via Ollama (gpt-oss, llama3, mistral, etc.) |
| Backend | FastAPI (Python 3.11+) |
| Frontend | React 18 + TypeScript + Vite |
| GPU | CUDA 12.8 with Flash Attention 2 |
| Package Manager | uv (10x faster than pip) |
# Windows
winget install Ollama.Ollama
# Pull a model
ollama pull gpt-ossdownload_model.batstart.batThat's it! The browser opens automatically at http://localhost:5173
| Mode | Description | VRAM Usage |
|---|---|---|
| Performance | Both models stay in VRAM. Fastest. | 16+ GB |
| Batch | OCR all files first, then parse all. Good for multiple files. | 12+ GB |
| Low VRAM | Unload OCR after each file. Slowest but works on 8GB. | 8 GB |
| Preset | Resolution | Speed | Best For |
|---|---|---|---|
| Tiny | 512px | ~8s/page | Quick previews, simple docs |
| Small | 640px | ~12s/page | Standard documents |
| Base | 1024px | ~25s/page | Most documents (default) |
| Large | 1280px | ~40s/page | Dense text, fine print |
VisionIngest/
├── backend/ # FastAPI application
│ ├── templates/ # Document templates (JSON schemas + prompts)
│ ├── ocr_adapter.py # DeepSeek-OCR-2 adapter
│ ├── llm_parser.py # LLM parser adapter
│ └── main.py # API endpoints
├── frontend/ # React UI
│ └── src/
│ ├── components/ # UI components
│ └── App.tsx # Main app
├── models/ # AI models (downloaded separately)
├── assets/ # Images for README
├── start.bat # One-click startup script
└── requirements.txt # Python dependencies
Business Source License 1.1
- Non-commercial use only
- Converts to Apache 2.0 on 2029-01-01
See LICENSE for full terms.