Skip to main content
JCore is a voice assistant that remembers. Its call-name is JIA. You talk to it through a microphone in the room, and it answers through a speaker. Unlike a stateless assistant, it keeps what it learns in a memory store, a vector database and a knowledge graph, so later requests can draw on earlier ones. The whole pipeline is designed to run on hardware you own: wake word, speech recognition, language model and speech synthesis all run locally.
JCore is in development. The components below are the prototyping stack. Source code, setup instructions and the hardware build report aren’t published yet, so this guide describes the design rather than how to install it.

​
What it does

​
The layers

JCore is organized into four layers, from hearing a request to acting on it, on top of a shared infrastructure layer.

Perception

Microphone, wake-word detection, speech-to-text and text-to-speech.

Context & understanding

Memory, a vector database, a graph database and a knowledge graph.

Reasoning

Intent parsing, the language model, and agents and pipelines.

Execution & control

Speaker output, Home Assistant as the IoT hub, and APIs.
A local network, a storage module and a dialog management module connect the layers. See Architecture for the component in each slot.

​
Hardware

These are the boards and devices used for development and prototyping:

​
Next steps

Architecture

Every component in the stack, layer by layer, and the infrastructure it runs on.