System Architecture

← Developer Docs · 简体中文

1. System Overview

AIDeepIn is an AI assistant platform with chat, knowledge bases (RAG), workflows and image generation.

┌─────────────────────────────────────────────────────────────────────────────┐
│                            User Layer                                       │
├──────────────────────────┬──────────────────────────────────────────────────┤
│   langchain4j-aideepin-web    │         langchain4j-aideepin-admin          │
│      (User Frontend Vue3)     │         (Admin Frontend Vue3)               │
│   • AI Chat / Drawing         │         • User Management                   │
│   • Knowledge Base Q&A        │         • Model Configuration               │
│   • Workflow Usage            │         • System Settings                   │
└──────────────────────────┴──────────────────────────────────────────────────┘
                                       │
                                       ▼
┌─────────────────────────────────────────────────────────────────────────────┐
│                      API Gateway Layer (Spring Boot)                        │
├──────────────────────────┬──────────────────────────────────────────────────┤
│        adi-chat               │              adi-admin                       │
│     (User API endpoints)      │         (Admin API endpoints)                │
└──────────────────────────┴──────────────────────────────────────────────────┘
                                       │
                                       ▼
┌─────────────────────────────────────────────────────────────────────────────┐
│                      Core Business Layer (adi-common)                       │
├──────────────┬──────────────┬──────────────┬──────────────┬─────────────────┤
│  Chat Service │  RAG Service │  Workflow    │ Image Gen    │  MCP Service    │
│               │              │  Engine      │ Service      │                 │
├──────────────┴──────────────┴──────────────┴──────────────┴─────────────────┤
│                    Model Service Layer (languagemodel)                      │
├─────────────┬─────────────┬─────────────┬─────────────┬────────────────────┤
│   OpenAI    │  DeepSeek   │  DashScope  │ SiliconFlow │     Ollama         │
└─────────────┴─────────────┴─────────────┴─────────────┴────────────────────┘
                                       │
                                       ▼
┌─────────────────────────────────────────────────────────────────────────────┐
│                         Data Storage Layer                                   │
├─────────────────────┬─────────────────────┬─────────────────────────────────┤
│   PostgreSQL        │       Redis         │       Neo4j (optional)          │
│   (pgvector+AGE)    │      (cache)         │       (graph database)         │
└─────────────────────┴─────────────────────┴─────────────────────────────────┘

2. Business Modules

2.1 AI Chat

The core module of the system, supporting multi-turn conversations with AI.

Chat Flow:

User input (text/voice/image)
    ↓
Pre-processing (voice-to-text, quota validation)
    ↓
Context enhancement (short-term/long-term memory + linked knowledge base → RAG retrieval / linked MCP → tool calls)
    ↓
LLM generates answer (streaming output)
    ↓
Post-processing (TTS synthesis, token billing, save history)

Core Capabilities:

Multi-model switching: Users can select different AI models in a conversation
Multi-turn memory: Short-term memory (current session context) + Long-term memory (key information across sessions)
Multimodal input: Text, images, voice
TTS synthesis: Convert AI responses to speech output
RAG-enhanced responses: When linked to a knowledge base, AI generates answers based on retrieved content
MCP tool calling: External tools can be invoked during conversations (e.g. search engines, database queries)
Multi-answer comparison: Multiple AI responses can be generated for the same question

2.2 Knowledge Base (RAG)

After importing documents into a knowledge base, AI chat can generate more accurate answers based on the knowledge base content.

RAG Architecture:

┌─────────────┐    ┌─────────────┐    ┌─────────────┐
│   Document  │ -> │   Document  │ -> │  Text       │
│   Upload    │    │   Parsing   │    │  Chunking   │
└─────────────┘    └─────────────┘    └─────────────┘
                                             │
                                             ▼
┌─────────────┐    ┌─────────────┐    ┌─────────────┐
│   Vector    │ <- │  Embedding  │ <- │  Text       │
│   Storage   │    │             │    │  Chunks     │
└─────────────┘    └─────────────┘    └─────────────┘
       │
       ▼
┌─────────────┐    ┌─────────────┐    ┌─────────────┐
│   User      │ -> │  Similarity │ -> │   Context   │
│   Question  │    │  Search     │    │  Building   │
└─────────────┘    └─────────────┘    └─────────────┘
                                             │
                                             ▼
                                      ┌─────────────┐
                                      │  LLM        │
                                      │  Generation │
                                      └─────────────┘

Document Processing Flow:

Upload documents (PDF/Word/TXT etc.)
    ↓
Parse document content, split into segments
    ↓
Index (optional: vector index, graph index, or both)
    ↓
User asks a question
    ↓
Retrieve relevant segments (vector similarity search / graph relationship query)
    ↓
Inject retrieved results into LLM prompt, generate answer

Index Types:

Vector Index: Converts text to vectors and retrieves via semantic similarity. Stored in pgvector
Graph Index: Extracts entities and relationships from documents to build a knowledge graph. Stored in Apache AGE
Hybrid Retrieval: Uses both vector and graph retrieval with combined ranking

Knowledge Base Configuration:

Public/Private: Public knowledge bases can be used by other users
Retrieval parameters: Max recall count, minimum similarity threshold
Chunking parameters: Document segment size, overlap length

2.3 Workflow

Build automated AI processing pipelines through visual orchestration of processing nodes.

Node Types:

Node	Purpose
Start Node	Workflow entry point, receives user input
Classification Node	Classifies input intent, determines the next branch
Knowledge Retrieval Node	Searches linked knowledge bases
LLM Answer Node	Calls an AI model to generate a response
HTTP Request Node	Calls an external API
Template Node	Generates structured output from a template
Switch/Conditional Node	Selects different branches based on conditions
Keyword Extraction Node	Extracts keywords from text
FAQ Extraction Node	Extracts Q&A pairs from text
Human Feedback Node	Pauses for human confirmation
Image Generation Node	Generates images
Email Node	Sends email

The table above is an architectural summary; the workflow canvas currently offers 17 node types — see the User Guide · Apps & Workflows.

Workflow and Chat Relationship: Workflows can internally search knowledge bases, call MCP tools and generate answers; users run workflows from the "Apps" page or trigger them via the open API.

2.4 MCP (Model Context Protocol)

MCP is a standardized protocol for integrating external tools, allowing AI to call external services during conversations.

Supported Transport Methods:

SSE (Server-Sent Events)
Stdio (standard I/O)
Docker containers
Remote services
WASM (WebAssembly)

Usage:

Administrators configure MCP services and their parameters
Users enable MCP services in conversations
AI automatically calls MCP tools as needed

2.5 Image Generation

Generate images from text descriptions:

User enters description → Select image model → Generate image → Save to gallery

Supports multiple image generation platforms (OpenAI, DashScope, SiliconFlow, etc.)
Generated images can be set as public (with watermark) or private
Supports likes and comments

3. Module Relationships

                              ┌──────────┐
                              │ User Web │
                              └────┬─────┘
                                   │ HTTP / SSE
                              ┌────▼─────┐
                              │ Backend  │
                              └────┬─────┘
                                   │
     ┌─────────┬─────────┬─────────┼─────────┬─────────┐
     │         │         │         │         │         │
┌────▼───┐┌───▼────┐┌───▼────┐┌──▼────┐┌───▼─────┐
│  Chat  ││ Image  ││Workflow││Knowledge││   MCP   │
│        ││Generation││      ││  Base  ││         │
└────┬───┘└───┬────┘└──┬────┘└───┬────┘└───┬─────┘
     │        │         │         │         │
     └────────┴────┬────┴─────────┴────┬────┘
                   │                   │
              ┌────▼────┐         ┌────▼────┐
              │LLM      │         │RAG      │
              │Services │         │Retrieval│
              │(multi-  │         │(vector+ │
              │platform)│         │graph)   │
              └─────────┘         └─────────┘

Interaction Relationships:

Chat ↔ Knowledge Base: Conversations can link to knowledge bases for RAG-enhanced responses
Chat ↔ MCP: Conversations can invoke MCP external tools
Workflow ↔ Knowledge Base: Workflow nodes can search knowledge bases
Workflow ↔ MCP: Workflow nodes can call MCP services
Workflow ↔ Image Generation: Workflows include image generation nodes
Workflow ↔ Chat: Workflow nodes can call LLM to generate answers

4. Users & Permissions

User Model:

Regular users: Register via email, use system features
Administrators: Manage system configuration, users, models, etc. via the admin panel

Quota Controls:

Daily token usage limits
Request rate limits
Image generation quantity limits

Privacy Controls:

Conversations, knowledge bases, workflows, and images can all be set as public or private
Public content can be viewed and used by other users

5. Admin Features

Feature	Description
Dashboard	System overview and monitoring
User Management	Create, edit users, set quotas
Conversation Management	View all user conversations, manage preset conversation templates
Knowledge Base Management	View and manage all knowledge bases
Model Management	Configure model platforms and AI models
MCP Management	Configure MCP services
Workflow Management	Manage workflow components
System Settings	Storage config, TTS/ASR config, quota config, rate limiting

Previous: Developer Docs contents · Next: Model Platform Integration Guide

System Architecture ​

1. System Overview ​

2. Business Modules ​

2.1 AI Chat ​

2.2 Knowledge Base (RAG) ​

2.3 Workflow ​

2.4 MCP (Model Context Protocol) ​

2.5 Image Generation ​

3. Module Relationships ​

4. Users & Permissions ​

5. Admin Features ​