SentrySearch is a semantic search tool for video footage that lets you find and extract clips by typing natural language descriptions of what you're looking for.
The tool works by splitting videos into overlapping chunks and embedding each chunk using either Google's Gemini Embedding API, Alibaba's Qwen3-VL model via DashScope, or a local Qwen3-VL model. These embeddings are stored in a ChromaDB database. When you search with text or an image query, it embeds your input into the same vector space, finds the best matching video chunk, and automatically trims and saves that section as a clip. The README highlights that the local model backend is fast because it avoids repeated API calls.
You should consider this tool if you work with dashcam footage, security video, or other video archives where you need to locate specific events quickly. It suits projects where you have multiple video files and want semantic search without manually reviewing hours of footage. The tool supports three embedding backends: Gemini Embedding 2 (requires an API key), Qwen3-VL via Alibaba DashScope (requires a separate API key), or a local Qwen3-VL model (no API key needed). The README positions this as the first step in a pipeline, with companion tools available for stitching multi-camera footage and redacting sensitive information. Installation requires Python 3.11 or 3.12 and ffmpeg for video processing.
Development activity shows consistent maintenance with bug fixes and feature additions. The project includes a codebase walkthrough video to help users understand the implementation. The maintainer has explicitly stated that the GitHub repository is the only official source and warns against using mirrors or republished versions.
