Enchanted
Introduction
Features and overview
Enchanted is a personal AI interface focused on privacy. Its goal is to leverage the most advanced models while ensuring that user identity and data has every possible protection. Design decisions in Enchanted prioritize the ability to access the best capabilities while being at the privacy pareto frontier. While the official release of Enchanted is opinionated about features, models, and integrations, developers can use the open-source version of Enchanted for greater flexibility.
Models
Enchanted relies on a variety of models.Completions Model
Use open-source models (Ollama, vLLM, llamacpp) or closed-source models
compatible with the OpenAI API. Completions model is used for chats, memory
and handling the agent.
Embeddings Model
Choose from open-source (Ollama), closed-source, or local embeddings.
Enchanted comes packaged with a open soruce, local embeddings model (JinaAI)
running on the ONNX inference engine. Embeddings model is responsible for
the agent memory and semantic search.
Anonymizer Model
The Anonymizer enables you to use advanced models
like GPT-3 or GPT-4.1 by performing semantic replacements on requests using
a local model trained to understand
PII.
STT
Any model compatible with the OpenAI API can be used for speech-to-text. The
official release of Enchanted uses Whisper running in a trusted execution
environment.
TTS
Any model compatible with the OpenAI API can be used for text-to-speech. The
official release of Enchanted uses the Kokoro model running in a trusted
execution environment.