SentrySearch

SentrySearch is an open-source tool designed to solve the pain of finding specific moments in long videos—think dashcam logs, vlog dumps, or surveillance archives. You type a natural language query like “a red truck running a red light,” and it scans all raw MP4 files to pinpoint the most semantically relevant segment, then auto-trims and exports it as a standalone clip. The entire workflow skips transcription, manual tagging, or timeline scrubbing.

Under the hood, it leverages Google’s Gemini Embedding 2, a native multimodal model that directly embeds raw pixels, motion, temporal context, and audio without intermediate text conversion. This means fewer information losses compared to transcription-first pipelines, especially in purely visual scenes or videos with subtle actions. It also offers an image-search mode: feed a reference screenshot and it returns visually similar segments, which is useful for finding matching compositions or stylistic references.

Drawbacks and caveats: the image-based search relies on overall visual similarity, not exact object matching—a red sedan might bring up other red cars with a similar shape.