volotat/Anagnorisis

What it solves

Anagnorisis is a local, private recommendation system designed to help users manage and filter their personal data (text, audio, images, and video) without relying on cloud services. It solves the problem of "AI slop" and generic recommendations by allowing users to train a personal model that aligns specifically with their individual preferences and interests.

How it works

Users interact with the system by rating their own data on a scale of 0 to 10. The system then fine-tunes a model to predict these preferences, which can then be used to sort and rank new data.

For discovery, it employs three search modes:

• Filename-based: Fuzzy matching of file names.
• Content-based: Uses embedding models to compare the file itself with a query.
• Metadata-based: Embeds text descriptions, including automatic descriptions generated by a descriptor model, zero-shot tags, and internal metadata (like EXIF or ID3).

Technologically, it uses a shared vector space for all content types via the jina-embeddings-v5-omni-small model and a specialized MiniCPM-o-4_5 descriptor model for natural-language file descriptions. The backend is built with Flask and PyTorch, while the frontend uses Bulma.

Who it’s for

It is for individuals who want a highly personalized, private information management system that runs entirely on their own hardware (requiring at least 8GB VRAM GPU or 16GB RAM for CPU mode) to avoid data leaks and cloud-based tracking.

Highlights

• Privacy-First: All data processing and storage happen locally on the user's computer.
• Multimodal Support: Handles images, music, text, and videos within a single shared vector space.
• Personalized Fine-Tuning: Users can actively train the model to "like" what they like through a rating system.
• Extensible Architecture: Supports external modules for new data types, such as YouTube or WebSearch.
• Local-Only Indexing: Remote files are never downloaded automatically, and searching is performed on the CPU to save GPU resources for background tasks.