System Architecture
XunLong employs a sophisticated multi-agent architecture orchestrated by LangGraph to deliver high-quality content generation.
Architecture Overview
Layer Breakdown
1. User Interface Layer
CLI Tool - Command-line interface providing:
- Simple, intuitive commands
- Real-time progress indicators
- Verbose mode for detailed logs
- Error handling and recovery
2. Agent Orchestration Layer
Coordinator - The master orchestrator that:
- Analyzes user requirements
- Decomposes complex tasks into subtasks
- Manages agent state and workflow
- Coordinates parallel execution
- Aggregates results
3. Core Agent Layer
Multiple specialized agents working in parallel:
- Search Agent - Executes web searches, extracts content, evaluates relevance
- Report Generator - Creates structured research reports with citations
- Fiction Generator - Writes creative stories with consistent characters and plot
- PPT Generator - Designs professional presentations with layouts and color schemes
- Iteration Agent - Handles content refinement requests
4. Support Service Layer
Supporting services for content processing:
- HTML Converter - Transforms Markdown to styled HTML
- Export Manager - Generates PDF, DOCX, PPTX files
- Storage Manager - Manages project files and versions
5. LLM Service Layer
Foundation layer powering all intelligence:
- LLM Manager - Manages multiple LLM providers, handles API calls
- Observability - Tracks all LLM interactions via LangFuse
Key Design Principles
Separation of Concerns
Each agent has a single, well-defined responsibility:
- Search agents only search
- Generation agents only generate
- Export managers only export
This modularity makes the system:
- Maintainable - Easy to update individual components
- Testable - Each agent can be tested in isolation
- Scalable - New agents can be added without affecting existing ones
State Machine Workflow
Built on LangGraph's state machine paradigm:
- Deterministic - Predictable execution flow
- Resumable - Can recover from failures
- Observable - Every state transition is tracked
Parallel Execution
XunLong maximizes throughput through parallelization:
- Multiple search queries execute simultaneously
- Independent content sections generate in parallel
- Export operations don't block generation
Technology Stack
| Layer | Technology |
|---|---|
| Workflow Orchestration | LangGraph |
| LLM Integration | LangChain |
| LLM Providers | OpenAI, Anthropic, DeepSeek |
| Web Automation | Playwright |
| PDF Generation | WeasyPrint |
| Office Export | python-pptx, python-docx |
| Observability | LangFuse |
Next Steps
- Dive deeper into the Multi-Agent System
- Understand the Workflow
- Explore LLM Integration