What it does SentrySearch chunks your video files, embeds each chunk via Gemini Embedding 2, Qwen3-VL (local), or Alibaba DashScope, and stores vectors in ChromaDB. Query with text or an image; it returns ranked timestamped matches and auto-trims the best one into a standalone clip. There’s also a highlights mode that surfaces statistically anomalous chunks when you don’t know what to search for.

The interesting bit The author treats video search as a solved problem by not solving video understanding at all — just throw overlapping chunks at a multimodal embedding API and let cosine similarity do the work. The clever part is the packaging: preprocessing (downscale to 480p, 5 fps), deduplication thresholds, and a three-project pipeline (search → merge multi-cam → auto-redact) that turns raw Tesla Sentry footage into usable evidence.

Key highlights

• Three backends: Gemini API (default, best quality), Alibaba DashScope (qwen-cloud), or fully local Qwen3-VL (auto-detects 2B vs 8B based on your GPU/RAM)
• Image search: drop a reference photo to find visually similar scenes
• highlights anomaly detection with three scoring methods (kNN, centroid, LOF) and query-relative anomaly modes
• Built-in deduplication (--dedupe 0.9) prevents the same event from flooding results across overlapping chunks
• Optional Tesla metadata overlay and companion tools for multi-cam stitching (SentryMerge) and face/license-plate redaction (SentryBlur)

Caveats

• Requires Python 3.11–3.12; PyTorch wheels don’t support 3.13+
• Local backend is impractical on Intel Macs or CPU-only machines (falls back to float32, “too slow and memory-hungry”)
• Anomaly detection isn’t magic: lens flares, night frames, and sensor glitches rank as “interesting” — use --exclude-baseline and --against to constrain it

Verdict Essential if you’re sitting on terabytes of dashcam or security footage and currently find events by scrubbing timelines. Skip if you don’t have a GPU or API budget and expect local inference to just work — the hardware requirements are real.