// the find
ssrajadh/sentrysearch
Semantic search over videos using Gemini Embedding 2 or Qwen3-VL.
SentrySearch does semantic search over local video footage using native video embeddings from Gemini Embedding 2, Alibaba's Qwen3-VL, or a local Qwen model — no transcription, no frame captioning. You describe what you're looking for in plain text, it returns a trimmed clip. Primary audience is Tesla dashcam owners, but it works on any MP4/MOV footage.
The core idea is technically sound: native video-to-vector embeddings mean you're actually comparing visual content against text queries in the same space, not doing OCR or captioning as a proxy. The local backend is well-thought-out — MRL dimension truncation to 768 dims, 5fps downscaling, still-frame skipping, and 4-bit quantization together make it actually usable on consumer hardware rather than a demo that needs an A100. The sibling tool ecosystem (SentryMerge for multi-cam stitching, SentryBlur for redaction) and the --last handoff pattern are a sensible way to chain operations without building a monolith. The dead-letter queue for failed chunks is good ops hygiene that most hobby projects skip.
Chunk boundary splitting is a real problem they acknowledge but haven't solved — if an incident spans two chunks, you may miss it or get a partial clip, and 'overlap helps but isn't perfect' is doing a lot of work there. The cost model for DashScope is documented as 'measure it yourself' while Gemini gets a concrete $/hr figure, which makes backend comparison difficult before you commit to indexing a large archive. The highlights anomaly detection is basically outlier scoring on embeddings, which means sensor glitches, exposure changes, and night-to-day transitions rank high alongside actually interesting events — the --exclude-baseline and --dedupe flags help but push the tuning burden onto the user. Still-frame detection via JPEG file size comparison is a heuristic that will quietly skip chunks with subtle motion (a slow pedestrian, gradual lane change) without telling you.