Architecture Overview

RiceBall is not just a tech stack, but a comprehensive AI Agent & Knowledge Base Platform.

Core Concepts

The diagram below shows the relationships between the core entities in the RiceBall system:

Assistant: The core entity for user interaction. It defines "Who am I" (System Prompt), "What brain do I use" (Model), "What do I remember" (Knowledge Base), and "What can I do" (Tools).
Channels: The interface for external communication.
- Adapter Pattern: Normalized handling of messages from different platforms (WeCom, Telegram, etc.).
- Identity Mapping: Automatically maps external user IDs (e.g., WeCom UserID) to internal system users, ensuring conversation isolation.
Model Provider & Model:
- Provider: Manages credentials for AI service providers (e.g., OpenAI API Key).
- Model: Specific model configuration (e.g., GPT-4o). RiceBall abstracts away the differences between providers.
RAG (Knowledge Base): Vector storage based on ChromaDB.
- Retrieval: Allows users to upload documents, and the Assistant automatically retrieves relevant snippets when answering, enabling "open-book exams".
- Memory (Write-Back): A unique feature that allows Assistants to write information back to the knowledge base, enabling long-term memory formation and note-taking.
Tools & MCP:
- Tools: Functional capabilities that the Assistant can call.
- MCP: Uses the Model Context Protocol to standardize the integration of tools and context, making it very easy to extend Assistant capabilities.

RiceBall adopts a modern separated frontend and backend architecture.