volotat/Anagnorisis

What it solves

Anagnorisis is a local, private recommendation system designed to replace cloud-based services that optimize for engagement rather than user preference. It allows users to build a personalized model of their interests by rating their own local or remote files (music, images, text, and video), ensuring that personal data and preference models remain entirely on the own machine.

How it works

The system uses a feedback loop of explicit ratings (0-10) and model training. It processes files through three search modes: filename-based (fuzzy matching), content-based (embedding the file itself), and metadata-based (embedding text descriptions, tags, and sidecar .meta files).

To create recommendations, the system generates "memory files" (Markdown) that act as text proxies for rated files, containing descriptions, tags, and metadata. A recommendation model is then trained on these pairs of descriptions and ratings. This model predicts ratings for unrated files, which the user can then correct, further refining the model in a subsequent training round.

Who it’s for

Users who want a private, self-hosted recommendation engine for their personal media libraries. It is targeted at those with a PC or home server equipped with at least 8GB of VRAM GPU.

Highlights

• Privacy-First: All processing, storage, and training happen locally on the user's machine.
• Distributed Semantic Indexing: Uses .meta sidecar files to allow semantic search and recommendations for remote files without downloading the actual content.
• Multimodal Support: Handles images, music, audio, and video files.
• Standalone Core: Includes anagnorisis_core, a Python library and CLI that can be used independently of the web UI.
• Soft Hashing: Identifies files by content fingerprints rather than paths, ensuring ratings are preserved even if files are moved or renamed.