Gemini Can Now Natively Embed Video: Developer Builds Sub-Second Video Search
Google's Gemini Embedding 2 model can now natively embed raw video directly into a 768-dimensional vector space alongside text, enabling sub-second semantic search over hours of footage without transcription or frame captioning.
A developer has built SentrySearch, a CLI tool that indexes dashcam footage into ChromaDB and allows natural language queries like "green car cutting me off" to instantly find and trim matching video clips. The system works by projecting raw video pixels into the same vector space as text queries, making direct comparisons possible at the vector level.
How It Works
Native Video Embedding: Gemini Embedding 2 projects raw video directly into vector space
No Intermediate Text: No transcription, no frame captioning, no text middleman
Sub-Second Search: Query hours of footage in under a second
Automatic Trimming: Matched clips are automatically trimmed from source video
Technical Details
The tool splits videos into overlapping 30-second chunks, embeds each chunk using Gemini's API, and stores vectors in a local ChromaDB database. When searching, text queries are embedded into the same space and matched against stored video embeddings.
Indexing costs approximately $2.50 per hour of footage, with optimizations like still-frame detection that skips idle chunks (useful for security camera/sentry mode footage).
Why This Matters for Builders
New Capability: First practical implementation of native video embedding
Cost-Effective: $2.50/hour indexing cost makes it accessible
Real-World Use Case: Dashcam/sentry mode footage search
No Complex Infrastructure: Works with local ChromaDB
Open Source: Available on GitHub for customization
Potential Applications
Security camera review and search
Content moderation for video platforms
Media asset management
Research video analysis
Personal video library organization
The project demonstrates how new AI capabilities can be quickly turned into practical tools, showing indie developers the path from research paper to working product.
Source: GitHub - SentrySearch, Hacker News Discussion
