Skip to main content

AI Analysis

ArgusAI uses advanced AI to analyze video events and generate natural language descriptions.

Multi-Provider Support​

Choose from multiple AI providers with automatic fallback:

ProviderModelBest For
OpenAIGPT-4o miniGeneral accuracy, low cost
xAI GrokGrok 2 VisionFast responses
AnthropicClaude 3 HaikuDetailed analysis
GoogleGemini FlashSpeed, multi-language

Analysis Modes​

Single Frame​

Traditional snapshot-based analysis. Fastest and lowest cost.

Multi-Frame​

Extract 3-5 key frames from video clips for better context. Balanced approach.

Video Native​

Send full video clips to providers that support it (OpenAI, Gemini). Highest accuracy.

Frame Extraction​

ArgusAI uses intelligent frame selection:

Similarity Filtering​

Removes nearly identical frames to focus on unique moments.

Motion Scoring​

Prioritizes frames with high activity using optical flow analysis.

Adaptive Sampling​

Combines similarity and motion scoring for optimal frame selection.

Context Enhancement​

AI prompts include contextual information:

  • Camera Name: Location context (e.g., "Front Door", "Driveway")
  • Time of Day: Temporal context (morning, afternoon, evening)
  • Date: Seasonal and schedule context

Example enhanced prompt:

This footage is from the Front Door camera at 7:15 AM on December 23, 2025.
Describe what you see in these security camera frames.

MCP Context System​

The Model Context Provider (MCP) enriches AI descriptions with historical and entity data:

  • Entity Context: Information about recognized people and vehicles
  • VIP/Blocked Status: AI knows about VIP and blocked entities for personalized descriptions
  • Recent Activity: What the entity has done recently at this location
  • Pattern Extraction: Common activity patterns using TF-IDF algorithm
  • Entity Adjustments: Your corrections to entity names improve future descriptions

Performance Optimizations​

The MCP context system is optimized for speed:

  • Parallel Queries: All context data is gathered simultaneously
  • 80ms Timeout: Fail-open behavior ensures descriptions are never delayed
  • Smart Caching: Recent context is cached to reduce database queries

Context Metrics​

Monitor context system performance via:

  • API: GET /api/v1/ai/context-metrics
  • Returns cache hit rate, timeout count, and system statistics

Confidence Scoring​

Each AI description includes a confidence score (0-100):

  • 90-100: High confidence, clear detection
  • 70-89: Good confidence, likely accurate
  • 50-69: Moderate confidence, may need verification
  • Below 50: Low confidence, consider re-analysis

Custom Prompts​

Customize the AI prompt in Settings > AI Models:

  1. Edit the Analysis Prompt text area
  2. Use the Refine Prompt button for AI-assisted improvements
  3. Save changes

Prompt Variables​

Available placeholders:

  • {camera_name} - Camera friendly name
  • {timestamp} - Event time
  • {date} - Event date

Re-Analysis​

To re-analyze an event with updated settings:

  1. Open the event detail view
  2. Click Re-analyse
  3. Wait for new description
  4. Compare with previous description

Cost Tracking​

Monitor AI usage in Settings > AI Costs:

  • Per-provider token usage
  • Daily/monthly cost estimates
  • Cost cap configuration