Summary Anagnorisis is a local-first data management platform featuring a private, trainable recommendation engine. It addresses data privacy concerns by processing all user data locally, enabling personalized information filtering and recommendations. The system aims to create a user's "digital twin" for efficient navigation of online content.

How It Works Users rate local data (text, audio, images, video). Recommendation models fine-tune on these ratings. A recent shift (v0.3.4+) uses a unified omni-descriptor model (e.g., MiniCPM-o-4_5) to convert all modalities into text descriptions. Search and recommendations operate on these text embeddings, promoting consistent, scalable cross-modal experiences and simplifying future development towards a single preference model.

Quick Start & Requirements Docker is the preferred installation. Clone the repo, configure docker-compose.override.yaml with absolute paths for data/config, and run docker compose up -d. Prerequisites: Docker, NVIDIA Container Toolkit/WSL2 for GPU. Recommended: Nvidia GPU (8GB VRAM), 32GB RAM. Container storage: ~45GB.

Highlighted Details

• Local-First Data Privacy: All data processing and storage are strictly local.
• Trainable Recommendation Engine: Models fine-tune on user ratings for personalized outputs.
• Unified Omni-Descriptor Model: Recent focus on a single model for cross-modal text embedding, unifying search and recommendations.
• Multi-Instance Support: Enables running multiple independent Anagnorisis instances.

Maintenance & Community Details on active maintenance, core contributors, or community channels (Discord, Slack) are absent. The project appears individually or small-team maintained, with a roadmap accessible via its wiki.

Licensing & Compatibility The specific open-source license is not stated. The platform defaults to localhost access only, with warnings against internet exposure due to incomplete security measures.

Limitations & Caveats The ongoing architectural shift to the omni-descriptor model may introduce performance regressions (slower, more demanding) and initial accuracy issues. Significant storage (~45GB) and hardware (recommended 8GB VRAM GPU, 32GB RAM) are required. Security for external access is incomplete. Initial setup involves lengthy model downloads and data caching. The license remains unspecified.