Customization Guide¶

Prompt Management¶

Export Current Prompts¶

# Export all prompts to YAML
netintel-ocr config export-prompts prompts.yaml

# Export specific category
netintel-ocr config export-prompts network-prompts.yaml --category network

Import Custom Prompts¶

# Import modified prompts
netintel-ocr config import-prompts custom-prompts.yaml

# Validate prompts
netintel-ocr config validate-prompts custom-prompts.yaml

Prompt Structure¶

network_detection:
  description: "Detect network diagrams in images"
  prompt: |
    Analyze this image for network architecture elements:
    - Routers, switches, firewalls
    - Network zones and segments
    - Connection types and protocols

    Return confidence score 0-1.

component_extraction:
  description: "Extract network components"
  prompt: |
    Identify all network components:
    {
      "components": [...],
      "connections": [...],
      "zones": [...]
    }

Key Tuning Parameters¶

1. Detection Confidence Threshold¶

# Default: 0.7
netintel-ocr process file document.pdf --confidence-threshold 0.9

# Lower for more detection (may include false positives)
netintel-ocr process file document.pdf --confidence-threshold 0.5

2. Model Temperature¶

# In prompts.yaml
network_extraction:
  temperature: 0.3  # Lower = more deterministic
  max_tokens: 4096
  top_p: 0.9

3. Context Window Size¶

# Surrounding text for context extraction
netintel-ocr --context-before 1000 --context-after 1000 document.pdf

# Disable context for speed
netintel-ocr process file document.pdf --no-context

4. Extraction Strategies¶

# Fast extraction (less accurate)
netintel-ocr process file document.pdf --fast-extraction

# Comprehensive extraction (slower)
netintel-ocr process file document.pdf --comprehensive

# Multi-pass extraction
netintel-ocr --multi-pass 3 document.pdf

5. Output Verbosity¶

# Minimal output
netintel-ocr process file document.pdf --output minimal

# Include all metadata
netintel-ocr process file document.pdf --output detailed

# Custom fields
netintel-ocr db query --fields "components,security,recommendations" document.pdf

6. Mermaid Syntax Preferences¶

# In config.yaml
mermaid:
  direction: TB  # Top-Bottom (or LR, RL, BT)
  theme: default
  node_shape: rectangle  # Or circle, diamond, hexagon
  include_labels: true
  simplify_connections: false

Model Selection Tips¶

Vision Models Comparison¶

Model	Speed	Accuracy	Memory	Best For
qwen2.5vl:7b	Fast	High	8GB	General use
llava:13b	Medium	Very High	16GB	Complex diagrams
minicpm-v	Very Fast	Medium	4GB	Quick processing
cogvlm	Slow	Highest	32GB	Critical accuracy

Selecting Models by Document Type¶

# Technical specifications
netintel-ocr --model Nanonets-OCR-s:latest \
             --network-model cogvlm:latest \
             technical-spec.pdf

# Marketing materials (simpler diagrams)
netintel-ocr --model minicpm-v:latest \
             --network-model minicpm-v:latest \
             marketing-doc.pdf

# Security documentation
netintel-ocr --model qwen2.5vl:7b \
             --security-focus \
             security-guide.pdf

Custom Processing Pipelines¶

Define Custom Pipeline¶

# custom_pipeline.py
from netintel_ocr import Pipeline

pipeline = Pipeline()
pipeline.add_step("ocr", model="Nanonets-OCR-s:latest")
pipeline.add_step("detect", confidence=0.8)
pipeline.add_step("extract", strategy="comprehensive")
pipeline.add_step("validate", fix_errors=True)
pipeline.add_step("context", window=2000)

Use Custom Pipeline¶

netintel-ocr config set pipeline custom_pipeline.py document.pdf

Advanced Prompt Engineering¶

Component Extraction Enhancement¶

component_extraction:
  prompt: |
    Extract network components with these specific details:
    1. Component type (router/switch/firewall/server/endpoint)
    2. Component name/label
    3. IP addresses or network ranges
    4. Security zone assignment
    5. Criticality level (high/medium/low)

    For each connection include:
    - Source and destination components
    - Protocol/port information
    - Direction (unidirectional/bidirectional)
    - Bandwidth or capacity if shown

Security Analysis Focus¶

security_analysis:
  prompt: |
    Analyze security aspects:
    - Identify trust boundaries and security zones
    - Detect exposed services and entry points
    - Map data flows crossing zone boundaries
    - Identify potential attack vectors
    - Suggest security improvements

    Priority: Accuracy over speed

Performance Optimization¶

Batch Processing¶

# Process multiple files efficiently
netintel-ocr process batch \
             --max-parallel 4 \
             --cache-models \
             *.pdf

GPU Acceleration¶

# Enable GPU
export CUDA_VISIBLE_DEVICES=0
netintel-ocr process file document.pdf --gpu

# Multi-GPU
export CUDA_VISIBLE_DEVICES=0,1
netintel-ocr --gpu --parallel-pages 2 document.pdf

Caching¶

# Enable result caching
netintel-ocr config set cache.dir /tmp/netintel-cache document.pdf

# Clear cache
netintel-ocr system clear-cache

Validation Rules¶

Custom Validation¶

validation_rules:
  network:
    - rule: "All components must have unique IDs"
    - rule: "Connections must reference existing components"
    - rule: "Security zones must be properly labeled"

  mermaid:
    - rule: "No orphaned nodes"
    - rule: "Valid Mermaid syntax"
    - rule: "Consistent node naming"

Apply Validation¶

netintel-ocr config set validation.rules custom-rules.yaml document.pdf

Additional Resources¶

Full Documentation: https://visionml.net/docs
PyPI Package: https://pypi.org/project/netintel-ocr/
Example Prompts: https://github.com/VisionMLNet/NetIntelOCR/tree/main/prompts

Next Steps¶

Vector Search Guide - Configure semantic search
Deployment Guide - Production setup
Quick Start Guide - Basic usage