SentrySearch ships Gemini video search with in-out timestamps and ffmpeg clip trimming

• SentrySearch has shipped as a Python CLI that uses Gemini Embedding 2 for native video embeddings, letting creators index footage without transcription and search it with text prompts through the project page.
• The tool already returns precise in/out timestamps and can auto-trim matching clips with ffmpeg, which makes it more like a selects assistant than a generic search demo, according to the HN launch thread.
• In the Hacker News discussion, editors immediately mapped it to rough-cut workflows: natural-language clip finding now, with EDL or Premiere-style export feeling like the obvious next step in the discussion roundup.

SentrySearch is positioned as semantic search for video footage, starting with dashcam video but clearly relevant to larger edit bins. Its project page says it splits footage into overlapping chunks, embeds those chunks with Gemini's native video model, stores them in local ChromaDB, and retrieves matches from plain-English queries like “red truck running a stop sign.”

The useful production detail is what happens after retrieval. According to the launch thread, the tool already outputs precise in/out timestamps and automatically trims clips with ffmpeg, so the result is not just a search hit but a usable excerpt for a selects reel or rough assembly.

The strongest creative signal came from the replies. One commenter described a future Premiere workflow where an editor could ask for every shot containing a specific subject and get an EDL back, while the creator replied that timestamped results already make an editable cut list a natural extension, as summarized in the discussion roundup.

That makes this less about surveillance-style search and more about replacing manual logging for repetitive review tasks. The current setup still has practical limits: the project page frames it as a CLI tool, requires a Gemini API key, and estimates indexing at about $2.50 per hour of footage.