1. System Architecture

1.1 OpenAI Agent Framework Integration

Based on OpenAI's AI Agent system architecture, ZaiXBT implements four core modules:

Planning Module

— Intent Recognition and Classification

"classification_cot": "Thinking steps:

1. Identify user input keywords and core intentions

2. Check completeness of required information

3. Determine if it matches specific classification features"

"classification_result": "chat/risk_analysis/drawing/investment"

— Task Decomposition

— Breaking down complex investment analysis into multi-round recursive queries

— Decomposing risk analysis tasks into address verification, data acquisition, risk assessment subtasks

Memory Module

— Multi-layer Memory Architecture

a. Short-term Memory (Redis)

— Latest N conversation records

— Real-time context maintenance

— Quick response support

b. Long-term Memory (Vector Database)

— Milvus-based vector storage

— Semantic similarity retrieval

— Cross-session knowledge extraction support

c. Memory Management Process

— Real-time conversation storage

— Vectorized storage and retrieval

— Relevance threshold filtering (score_threshold=0.6)

— Smart recall of Top-K relevant memories

— Context maintenance

— Maintaining ZaiXBT's personality traits

— Recording timestamp and dialogue sequence number for each user

— Preserving chat history from all entry points for permanent memory retention

Tools Module

Risk Analysis Tools:

async def analyze_address_risk(address):

"""Call BlockSec API for address risk analysis"""

• Implementation of asynchronous interaction with BlockSec API

Multi-Agent Market Analysis System

1. Recursive Problem Decomposition

— Initial problem analysis

— Core focus identification

— Multi-dimensional sub-problem breakdown

— Deep analysis coverage implementation

1. Multi-round Information Gathering

• First Round: Basic Data Collection

— Price trend analysis

— Trading volume evaluation

— Market sentiment scanning

• Second Round: Deep Mining

— On-chain data analysis

— Whale behavior tracking

— Related news interpretation

• Third Round: Comprehensive Verification

— Cross-validation of information

— Potential risk identification

— Trend change prediction

1. Report Generation Process

— Multi-dimensional data integration

— Risk factor analysis

— Trend judgment

— Investment advice generation

1. Output Optimization

— Formatted report generation

— Multiple export format support (MD/DOCX)

— Key information highlighting

1. Response Time Optimization

— Parallel data acquisition

— Asynchronous processing mechanism

— Cache strategy application

Action Module

— Response Generation

{
    "response_cot": "",    # Chain of thought
    "first_response": "",  # Initial answer
    "reflection": "",      # Reflection optimization
    "response": ""        # Final answer
}

— Formatted Output

— JSON structured response

— Maintaining role consistency