• QUICK-START-NVIDIA.md - NVIDIA GPU users (most common) - 30 min setup
• QUICK-START-AMD.md - AMD GPU users - 30 min setup

• NVIDIA-SETUP.md - Complete NVIDIA CUDA setup guide
• GPU-SETUP.md - AMD Vulkan setup guide (original)
• AGENTS.md - AI agent implementation details and requirements

• QUICKREF.md - Quick command reference
• SETUP.md - General setup documentation

Have an NVIDIA GPU? (RTX 20/30/40 series, Tesla, etc.)

# Check what you need to install
.\setup-nvidia.ps1

# Then follow: QUICK-START-NVIDIA.md

Have an AMD GPU? (Radeon RX 6000/7000 series)

# Use AMD Vulkan setup
# See: GPU-SETUP.md

# NVIDIA GPU users:
.\start-gpu-server-nvidia.ps1    # Terminal 1: GPU server
docker compose up                 # Terminal 2: Other services

# AMD GPU users:
.\start-gpu-server.ps1            # Terminal 1: GPU server  
docker compose up                 # Terminal 2: Other services

# Test everything
.\test-classification.ps1

# Stop everything
docker compose down
# Then Ctrl+C in GPU terminal

End-to-end document classification system with:

• Pre-processing checks (legibility, page/image count)
• Extraction via Docling + OCR for documents with images
• Guard rules and PII detection with redaction
• Local GPU-accelerated LLM (AMD Vulkan) for classification
• Verifier for second-pass validation
• Citation-based evidence for audit trails
• Web UI for easy document upload and results
• Safety monitoring for unsafe content detection

Web UI (port: file://) 
    ↓
Classification Server (port: 5055)
    ↓
GPU Server - llama.cpp (port: 8080) + Docling (port: 7000)

• Prompts: server/src/config/prompts.json (classifier/verifier and class rules).
• Thresholds: ROUTE_LOW, AUTO_ACCEPT in .env.
• Verifier engine: VERIFIER_ENGINE=openai|llama. For llama.cpp, set LLAMA_URL.
• Docling REST (optional): set DOCLING_URL if you have a Docling server.
• PII Redaction: REDACT_PII=true|false.
• Cross-verify with two LLMs: CROSS_VERIFY=true (requires both engines configured).
• Offline mode: OFFLINE_MODE=true to avoid any external network.

Local classifier

• Optional linear TF-IDF model can be placed at server/models/tfidf_svm.json.
• If missing, a heuristic classifier based on guard signals is used.

Front end

• Static React (CDN) in web/. No build tooling required.
• Accepts PDF and common image formats; displays safety, PII, policy, and status updates.

Notes

• For robust citations (page/bbox), integrate Docling or PDF engines with positional data in extractor/doclingAdapter.js and adjust citations.js.
• For multimodal verification with GGUF, run llama.cpp server locally and set VERIFIER_ENGINE=llama.
• Batch API: POST /api/process-batch with multipart files[] or JSON { paths: [...] } returns per-file results.