SentrySearch: Semantic Video Search with AI
Discover SentrySearch, the open-source tool that transforms hours of video footage into searchable clips using Google's Gemini Embedding API or local Qwen3-VL models. Just type 'red truck running a stop sign' and get perfectly trimmed video clips back. Perfect for Tesla dashcam analysis, security footage, or any MP4/MOV files. Features local GPU acceleration, Tesla metadata overlays, and automatic still-frame skipping to save costs and time.
SentrySearch: Revolutionize Video Footage Analysis with Semantic Search
Finding specific moments in hours of video footage just got dramatically easier. SentrySearch is an open-source Python tool that uses cutting-edge video embedding models to make your videos searchable by natural language queries.
How It Works (Pure Magic)
Instead of manually scrubbing through footage, SentrySearch:
- Chunks your videos into overlapping 30-second segments
- Embeds each chunk using Google's Gemini Embedding API or local Qwen3-VL models
- Stores embeddings in a local ChromaDB vector database
- Matches your text query ("car running red light") against video embeddings
- Automatically trims and saves the best matching clip
No transcription. No frame captioning. Direct video-to-text semantic matching.
🚀 Quick Start (5 Minutes)
# Install with uv (fastest package manager)
curl -LsSf https://astral.sh/uv/install.sh | sh
git clone https://github.com/ssrajadh/sentrysearch.git
cd sentrysearch
uv tool install .
# Setup API key (free tier available)
sentrysearch init
# Index your footage
sentrysearch index /path/to/videos
# Search and get clips!
sentrysearch search "red truck running stop sign"
Result: ./match_front_2024-01-15_14-30_02m15s-02m45s.mp4 ✅
🌐 Two Backends: Cloud or Local
| Backend | Pros | Cons | Cost |
|---|---|---|---|
| Gemini API (default) | Best quality, no GPU needed | Requires API key | ~$2.84/hour indexed |
| Local Qwen3-VL | Free, private, offline | Needs GPU/M1+ Mac | Free |
Local model auto-scales to your hardware:
- NVIDIA 18GB+ VRAM: Qwen8B (full precision)
- 8-16GB VRAM: 4-bit quantized (~6GB)
- M1/M2 Mac 16GB+: Qwen2B (~6GB)
Tesla Dashcam Superpowers
Got Tesla Sentry Mode footage? SentrySearch extracts embedded telemetry and burns HUD overlays:
sentrysearch search "car cutting me off" --overlay
Shows speed, GPS location, timestamp, even reverse-geocoded street names.
💰 Cost Optimization
- Still-frame skipping: Skips parked car footage automatically
- Preprocessing: 95% size reduction (480p@5fps chunks)
- Tunable chunks:
--chunk-duration 60 --overlap 5halves API costs
1 hour indexing = ~$2.84 (3,600 frames × $0.00079)
🛠️ Pro Features
# Save top 5 matches
sentrysearch search "pedestrian" --save-top 5
# Custom confidence threshold
sentrysearch search "event" --threshold 0.7
# Manage index
sentrysearch stats
sentrysearch remove old_footage/
sentrysearch reset
🎯 Use Cases
- Tesla owners: Analyze Sentry Mode incidents
- Security teams: Search surveillance footage
- Content creators: Find specific clips in raw footage
- Researchers: Query behavioral patterns in video datasets
- Anyone with dashcam/action cam footage
📦 Works With
✅ MP4, MOV files ✅ Tesla firmware 2025.44.25+ (HW3+) ✅ Regular dashcams, phone videos, security cams ✅ Python 3.11+, ffmpeg (auto-bundled)
🚀 Why SentrySearch Wins
- 2.6K GitHub stars, battle-tested
- MIT licensed, production-ready
- Sub-second search even on hours of footage
- Dual backend (cloud + local)
- Automatic trimming, no post-processing
- Cost-aware optimizations built-in
Stop fast-forwarding through footage. Start searching.