Ollama Local Setup

Run AI extraction locally with Ollama for complete data privacy.

When to Use Ollama

Data sovereignty requirements
Air-gapped environments
No recurring API costs
Privacy-sensitive documents

Prerequisites

Linux/macOS/Windows server
Minimum 16GB RAM (32GB recommended)
GPU recommended (NVIDIA with 8GB+ VRAM)
TaprNext admin access

Step 1: Install Ollama

Linux/macOS

curl -fsSL https://ollama.com/install.sh | sh

Windows

Download installer from ollama.com/download

Verify Installation

ollama --version
# ollama version 0.1.x

Step 2: Download Vision Model

For invoice extraction, you need a vision-capable model:

# Recommended: LLaVA (good balance)
ollama pull llava:13b

# Alternative: Smaller/faster
ollama pull llava:7b

# Alternative: Llama 3.2 Vision
ollama pull llama3.2-vision

Download takes 5-15 minutes depending on connection.

Step 3: Start Ollama Server

# Start server (runs on port 11434 by default)
ollama serve

# Or run as system service
sudo systemctl enable ollama
sudo systemctl start ollama

Verify server is running:

curl http://localhost:11434/api/tags
# Should list available models

Step 4: Configure TaprNext

Go to Tapr Settings
Select AI Provider tab
Choose Ollama (Local)
Enter server URL: http://localhost:11434 (or your server IP)
Select model: llava:13b
Click Save

Hardware Recommendations

Invoice Volume	Recommended Hardware
< 50/day	16GB RAM, CPU only (slow but works)
50-200/day	32GB RAM, NVIDIA RTX 3060 (12GB)
200-500/day	64GB RAM, NVIDIA RTX 4080 (16GB)
> 500/day	Multiple GPUs or cloud GPUs

Performance Comparison

Model	RAM	Speed	Accuracy
llava:7b	8GB	~5s/page	80-85%
llava:13b	16GB	~10s/page	85-90%
llama3.2-vision	12GB	~8s/page	85-88%

Troubleshooting

”Connection Refused”

Ensure Ollama server is running
Check firewall allows port 11434
Verify URL in TaprNext settings

”Model Not Found”

Run ollama list to see available models
Pull the model: ollama pull llava:13b

”Out of Memory”

Use smaller model (llava:7b)
Close other applications
Add more RAM or use GPU offloading

Slow Extraction

Consider GPU acceleration
Use smaller model for speed
Process invoices in off-hours