arama

Find similar images and videos — entirely on your machine.

Overview

arama is a desktop GUI application that uses offline AI to locate visually or aurally similar media files inside a chosen directory tree. There is no cloud service, no account, and no data leaves your device.

• Images are compared by CLIP visual embeddings (cosine similarity).
• Videos are compared by a weighted combination of CLIP frame embeddings and wav2vec2 audio embeddings.

Embeddings and thumbnails are cached in a local SQLite database so each directory only needs to be indexed once.

Why / When

arama works best with a reasonably modern desktop (an Apple Silicon Mac or a multi-core Linux/Windows machine). CPU-only inference is supported; a discrete GPU is not required.

Quick Start

Prerequisites

• Rust toolchain (stable, 2024 edition)
• An internet connection for the one-time model and ffmpeg download (a few hundred MB total)

Build and run

# Extract the source archive
tar xzf arama-vX.Y.Z.tar.gz
cd arama-vX.Y.Z

# Build and launch (release mode recommended for AI inference speed)
cargo run -p arama --release

The first launch opens a setup wizard that downloads:

• openai/clip-vit-base-patch32 — CLIP model for image similarity
• facebook/wav2vec2-base-960h — audio model for video similarity
• ffmpeg — frame / audio extraction for video files

All files are stored alongside the executable under .arama-local/. Once setup completes, no further network access is required.

Design Notes

• Offline-first. All AI inference runs locally with candle. No telemetry.
• iced GUI. Built on iced 0.14 with the snora shell framework. Side-nav pages: Explorer (directory tree + gallery tiling view), Cache (per-directory cache management), and Settings.
• localcache persistence. Embeddings and thumbnails are stored in a two-namespace SQLite database via localcache, keyed by file path. Re-indexing is triggered automatically when a file changes.
• Similarity threshold. Both image and video similarity default to 0.86 cosine similarity (dot product of unit-norm CLIP vectors).
• Supported formats. Images: png jpg jpeg webp gif bmp. Videos: mp4.

More Detail

Full documentation lives in docs/src/ and is structured for mdBook.