> ## Documentation Index
> Fetch the complete documentation index at: https://documentation.3cubed.vc/llms.txt
> Use this file to discover all available pages before exploring further.
# Architecture
> The components in each layer of the JIA stack, and the network and infrastructure they run on.
JCore's logical stack has four functional layers. Three cross-cutting pieces sit between them: a local network, a storage module and a dialog management module. Everything runs on one infrastructure layer.
## Perception
Perception turns sound into text and text back into sound.
| Component | Implementation |
| ------------------- | --------------------------------------------------------------------------------------------------------------------------- |
| Microphone | ESP32-S3 SoC |
| Wake-word detection | [OpenWakeWord](https://github.com/dscripka/openWakeWord), as a custom [Wyoming](https://github.com/rhasspy/wyoming) service |
| Speech-to-text | [Faster-Whisper](https://github.com/SYSTRAN/faster-whisper), on CTranslate2 |
| Text-to-speech | [Piper](https://github.com/rhasspy/piper), as a custom Wyoming service |
Wyoming is the protocol Home Assistant uses to talk to voice services. Running wake-word detection and speech synthesis as Wyoming services lets them plug into Home Assistant's voice pipeline.
## Local VA network
The voice-assistant devices sit on their own network.
| Component | Implementation |
| ---------------- | ------------------------------------------- |
| LAN access point | Raspberry Pi 4 Model B in access-point mode |
| VPN tunnel | [WireGuard](https://www.wireguard.com/) |
## Context & understanding
This layer is what gives JCore persistent recall. It keeps what the assistant has learned in several forms, each suited to a different kind of lookup.
| Component | Implementation |
| --------------------------------- | --------------------------------------- |
| Memory | [AnythingLLM](https://anythingllm.com/) |
| Vector database | [Chroma](https://www.trychroma.com/) |
| Graph database | [Neo4j](https://neo4j.com/) |
| Embedding graph / knowledge graph | Built on the stores above |
The **storage module** sits below this layer and connects it to reasoning.
## Reasoning
| Component | Implementation |
| ------------------ | ------------------------------------------------------------------- |
| Intent parser | Classifies what a request is asking for |
| LLM | [Mistral 7B v0.3](https://huggingface.co/mistralai/Mistral-7B-v0.3) |
| Agents & pipelines | Multi-step work that spans more than one model call or tool |
The **dialog management module** sits below this layer and connects it to execution.
## Execution & control
| Component | Implementation |
| --------- | ------------------------------------------------ |
| Speaker | ESP32-S3 SoC |
| IoT hub | [Home Assistant](https://www.home-assistant.io/) |
| APIs | [LangChain](https://www.langchain.com/) tools |
## Infrastructure
The server that runs the stack:
| Component | Part |
| --------- | ---------------------------------- |
| GPU | NVIDIA RTX 4000 SFF Ada Generation |
| CPU | Intel Core i5-13500 |
The prototyping boards, including a Jetson Orin Nano for local inference, are listed under [Hardware](/jcore#hardware) on the overview page.
## Not documented yet
* **How the modules talk to each other.** The storage and dialog management modules are named in the design, but their interfaces aren't written up.
* **Setup.** There are no install or configuration instructions yet.
* **The hardware build report.** It's planned but not published.