natural-language-video-search（自然语言视频搜索）

Natural Language Video Search

Search video files using natural language queries powered by Gemini Embedding 2's native video-to-vector embedding.

What This Skill Does

This skill lets you index video files (dashcam footage, security camera recordings, any mp4) into a local vector database, then search them by describing what you're looking for in plain English. The top match is automatically trimmed and saved as a clip.

For Tesla dashcam footage, an optional telemetry overlay can burn speed, GPS, location, and turn signal data onto trimmed clips.

Setup

Requires uv and Python 3.11+.

1. Clone and install:

git clone https://github.com/ssrajadh/sentrysearch.git
cd sentrysearch
uv sync

For Tesla overlay support (reverse geocoding):

uv sync --extra tesla

1. Set your Gemini API key:

sentrysearch init

This prompts for your key, writes it to .env, and validates it with a test embedding. You can also set GEMINI_API_KEY directly as an environment variable.

Commands

Index video files

sentrysearch index <directory_or_file>

Options: --chunk-duration (default 30s), --overlap (default 5s), --no-preprocess, --target-resolution, --target-fps, --skip-still / --no-skip-still, --verbose

Search indexed footage

sentrysearch search "<natural language query>"

Options: -n / --results (default 5), -o / --output-dir, --trim / --no-trim, --threshold (default 0.41), --overlay / --no-overlay (Tesla telemetry), --verbose

Apply Tesla telemetry overlay

sentrysearch overlay <video_file>
sentrysearch overlay <video_file> -o output.mp4

Burns a HUD overlay onto a Tesla dashcam video showing speed, GPS coordinates, location name, and turn signal status. Reads telemetry from SEI NAL units embedded in Tesla firmware 2025.44.25+. Also available as --overlay flag on the search command to automatically overlay the trimmed clip.

Check index stats

sentrysearch stats

How It Works

Video files are split into overlapping chunks. Still-frame detection can skip chunks with no meaningful visual change, eliminating unnecessary API calls — this is the primary cost saver for idle footage like sentry mode or security cameras. Chunks are also preprocessed (reduced frame rate and resolution) to shrink upload size and speed up transfers, though the Gemini API bills based on video duration at a fixed token rate, not file size, so preprocessing does not reduce per-chunk token cost. Each chunk is embedded as raw video using Gemini Embedding 2 (no transcription or captioning). Vectors are stored in a local ChromaDB database. Text queries are embedded into the same vector space and matched via cosine similarity. The top match is auto-trimmed from the original file via ffmpeg.

When To Use This Skill

• User asks to search through video files or footage
• User wants to find a specific moment in a video by describing it
• User asks to index or organize video footage for search
• User mentions dashcam, security camera, or surveillance clips
• User wants to find and extract a clip from a longer video
• User has Tesla dashcam footage and wants speed/GPS/location overlay on clips
• User wants to apply telemetry overlay to a Tesla video

Example Interactions

User: "Search my dashcam footage for a white truck cutting me off" Action: Run sentrysearch search "white truck cutting me off"

User: "Index all the video files in my Downloads folder" Action: Run sentrysearch index ~/Downloads

User: "Search for a red light and include the Tesla overlay on the clip" Action: Run sentrysearch search "running a red light" --overlay

User: "Add the speed and GPS overlay to this Tesla video" Action: Run sentrysearch overlay /path/to/tesla_video.mp4

User: "How much footage do I have indexed?" Action: Run sentrysearch stats

Rules

• Always run sentrysearch init or confirm GEMINI_API_KEY is set before indexing or searching.
• If ffmpeg is not found on PATH, the bundled imageio-ffmpeg fallback is used automatically.
• Indexing costs ~$2.84/hour of active footage with default settings. Cost is driven by the number of chunks sent to the API — footage with long idle periods (sentry mode, security cameras) will be significantly cheaper since still-frame skipping eliminates those chunks entirely. Warn the user before indexing large directories.
• Search results include similarity scores. Scores below the threshold (default 0.41) trigger a low-confidence prompt before trimming.
• The Tesla overlay requires firmware 2025.44.25+ for SEI metadata. Videos without Tesla metadata will skip the overlay gracefully.
• Requires Python 3.11+.