AgenticVision

Core vision library for AI agents — image capture, CLIP embedding, similarity search, and persistent visual memory.

What it does

AgenticVision gives AI agents persistent visual memory. Capture images, embed them with CLIP ViT-B/32 into 512-dimensional vectors, store them in a compact .avis binary format, and query them by similarity, time, or description.

Install

cargo install agentic-vision

Or add to your Cargo.toml:

[dependencies]
agentic-vision = "0.1"

Usage

use agentic_vision::{VisionStore, CaptureSource};

let mut store = VisionStore::open("observations.avis")?;

// Capture from file
let id = store.capture(
    CaptureSource::File("screenshot.png"),
    "Homepage after deploy"
)?;

// Find similar
let matches = store.similar(id, 5)?;
for m in matches {
    println!("  {} (similarity: {:.3})", m.description, m.score);
}

Key features

CLIP ViT-B/32 embeddings — 512-dimensional vectors via ONNX Runtime, with fallback mode when model is not present
Binary .avis format — 64-byte header, JSON payload, JPEG thumbnails. Single file, portable, no database
Similarity search — Brute-force cosine in 1-2 ms (top-5)
Visual diff — Pixel-level differencing with 8×8 grid region detection in <1 ms
Image capture — From files, base64, screenshots, or clipboard. Auto-resize and JPEG compression. Native screenshot support on macOS (screencapture) and Linux (gnome-screenshot/scrot/maim); clipboard capture via osascript (macOS) or xclip/wl-paste (Linux)

Performance

Operation	Time
Image capture (file → embed → store)	47 ms
Similarity search (top-5)	1-2 ms
Visual diff (pixel-level)	<1 ms
Storage per capture	~4.26 KB

MCP Server

For LLM integration via the Model Context Protocol, see agentic-vision-mcp.

License

MIT

agentic-vision 0.3.0