## β‘ Project Overview
**"BettaFish"** is an innovative multi-agent public opinion analysis system built from scratch. It helps break information cocoons, restore the original public sentiment, predict future trends, and assist decision-making. Users only need to raise analysis needs like chatting; the agents automatically analyze 30+ mainstream social platforms at home and abroad and millions of public comments.
> Betta is a small yet combative and beautiful fish, symbolizing "small but powerful, fearless of challenges".
See the system-generated research report on "Wuhan University Public Opinion": [In-depth Analysis Report on Wuhan University's Brand Reputation](./final_reports/final_report__20250827_131630.html)
See a complete system run example on "Wuhan University Public Opinion": [Video - In-depth Analysis Report on Wuhan University's Brand Reputation](https://www.bilibili.com/video/BV1TH1WBxEWN/?vd_source=da3512187e242ce17dceee4c537ec7a6#reply279744466833)
Beyond just report quality, compared to similar products, we have π six major advantages:
1. **AI-Driven Comprehensive Monitoring**: AI crawler clusters operate 24/7 non-stop, comprehensively covering 10+ key domestic and international social media platforms including Weibo, Xiaohongshu, TikTok, Kuaishou, etc. Not only capturing trending content in real-time, but also drilling down to massive user comments, letting you hear the most authentic and widespread public voice.
2. **Composite Analysis Engine Beyond LLM**: We not only rely on 5 types of professionally designed Agents, but also integrate middleware such as fine-tuned models and statistical models. Through multi-model collaborative work, we ensure the depth, accuracy, and multi-dimensional perspective of analysis results.
3. **Powerful Multimodal Capabilities**: Breaking through text and image limitations, capable of deep analysis of short video content from TikTok, Kuaishou, etc., and precisely extracting structured multimodal information cards such as weather, calendar, stocks from modern search engines, giving you comprehensive control over public opinion dynamics.
4. **Agent "Forum" Collaboration Mechanism**: Endowing different Agents with unique toolsets and thinking patterns, introducing a debate moderator model, conducting chain-of-thought collision and debate through the "forum" mechanism. This not only avoids the thinking limitations of single models and homogenization caused by communication, but also catalyzes higher-quality collective intelligence and decision support.
5. **Seamless Integration of Public and Private Domain Data**: The platform not only analyzes public opinion, but also provides high-security interfaces supporting seamless integration of your internal business databases with public opinion data. Breaking through data barriers, providing powerful analysis capabilities of "external trends + internal insights" for vertical businesses.
6. **Lightweight and Highly Extensible Framework**: Based on pure Python modular design, achieving lightweight, one-click deployment. Clear code structure allows developers to easily integrate custom models and business logic, enabling rapid platform expansion and deep customization.
**Starting with public opinion, but not limited to public opinion**. The goal of "WeiYu" is to become a simple and universal data analysis engine that drives all business scenarios.
> For example, you only need to simply modify the API parameters and prompts of the Agent toolset to transform it into a financial market analysis system.
>
> Here's a relatively active Linux.do project discussion thread: https://linux.do/t/topic/1009280
>
> Check out the comparison by a Linux.do fellow: [Open Source Project (BettaFish) vs manus|minimax|ChatGPT Comparison](https://linux.do/t/topic/1148040)
Say goodbye to traditional data dashboards. In "WeiYu", everything starts with a simple question - you just need to ask your analysis needs like a conversation
## πͺ Sponsors
LLM Model API Sponsor: Solomon Blog LionCC.ai; Programming Carpool codecodex.ai; Programming Computing Power VibeCodingAPI.ai:
1. Solomon Blog LionCC.ai has updated the "BettaFish WeiYu System - LionCC API Deployment Configuration Complete Guide" and is optimizing one-click deployment and cloud server invocation solutions.
2. VibeCodingapi.ai Lion Computing Platform has adapted to all LLM models of "BettaFish WeiYu System", including Claude Code, OpenAI Codex, and Gemini CLI programming development computing power. The quota price is 1:1 (100 yuan equals 100 USD quota).
3. Codecodex.ai Lion Programming Carpool System has achieved IP-free access to bypass Claude Code and OpenAI Codex restrictions. After following the official deployment tutorial, simply switch the BASE_URL invocation address and Token key invocation key to use the most powerful programming models.
Solomon LionCC BettaFish WeiYu Benefits: Open codecodex.ai Lion Programming Channel, scan the QR code to join the WeChat community, register for VibeCodingapi.ai Lion Computing, and receive 20 USD API quota (limited to the first 1,000 users).
Pay-as-you-go enterprise-grade AI resource platform offering a comprehensive set of AI models and APIs, plus multiple ready-to-use online AI apps:
302.AI is a pay-as-you-go enterprise AI resource hub that offers the latest and most comprehensive AI models and APIs on the market, along with a variety of ready-to-use online AI applications.
## ποΈ System Architecture
### Overall Architecture Diagram
**Insight Agent** Private Database Mining: AI agent for in-depth analysis of private public opinion databases
**Media Agent** Multimodal Content Analysis: AI agent with powerful multimodal capabilities
**Query Agent** Precise Information Search: AI agent with domestic and international web search capabilities
**Report Agent** Intelligent Report Generation: Multi-round report generation AI agent with built-in templates
### A Complete Analysis Workflow
| Step | Phase Name | Main Operations | Participating Components | Cycle Nature |
|------|------------|-----------------|-------------------------|--------------|
| 1 | User Query | Flask main application receives the query | Flask Main Application | - |
| 2 | Parallel Launch | Three Agents start working simultaneously | Query Agent, Media Agent, Insight Agent | - |
| 3 | Preliminary Analysis | Each Agent uses dedicated tools for overview search | Each Agent + Dedicated Toolsets | - |
| 4 | Strategy Formulation | Develop segmented research strategies based on preliminary results | Internal Decision Modules of Each Agent | - |
| 5-N | **Iterative Phase** | **Forum Collaboration + In-depth Research** | **ForumEngine + All Agents** | **Multi-round cycles** |
| 5.1 | In-depth Research | Each Agent conducts specialized search guided by forum host | Each Agent + Reflection Mechanisms + Forum Guidance | Each cycle |
| 5.2 | Forum Collaboration | ForumEngine monitors Agent communications and generates host summaries | ForumEngine + LLM Host | Each cycle |
| 5.3 | Communication Integration | Each Agent adjusts research directions based on discussions | Each Agent + forum_reader tool | Each cycle |
| N+1 | Result Integration | Report Agent collects all analysis results and forum content | Report Agent | - |
| N+2 | Report Generation | Dynamically select templates and styles, generate final reports through multiple rounds | Report Agent + Template Engine | - |
### Project Code Structure Tree
```
BettaFish/
βββ QueryEngine/ # Domestic and international news breadth search Agent
β βββ agent.py # Agent main logic
β βββ llms/ # LLM interface wrapper
β βββ nodes/ # Processing nodes
β βββ tools/ # Search tools
β βββ utils/ # Utility functions
β βββ ... # Other modules
βββ MediaEngine/ # Powerful multimodal understanding Agent
β βββ agent.py # Agent main logic
β βββ nodes/ # Processing nodes
β βββ llms/ # LLM interfaces
β βββ tools/ # Search tools
β βββ utils/ # Utility functions
β βββ ... # Other modules
βββ InsightEngine/ # Private database mining Agent
β βββ agent.py # Agent main logic
β βββ llms/ # LLM interface wrapper
β β βββ base.py # Unified OpenAI-compatible client
β βββ nodes/ # Processing nodes
β β βββ base_node.py # Base node class
β β βββ formatting_node.py # Formatting node
β β βββ report_structure_node.py # Report structure node
β β βββ search_node.py # Search node
β β βββ summary_node.py # Summary node
β βββ tools/ # Database query and analysis tools
β β βββ keyword_optimizer.py # Qwen keyword optimization middleware
β β βββ search.py # Database operation toolkit
β β βββ sentiment_analyzer.py # Sentiment analysis integration tool
β βββ state/ # State management
β β βββ __init__.py
β β βββ state.py # Agent state definition
β βββ prompts/ # Prompt templates
β β βββ __init__.py
β β βββ prompts.py # Various prompts
β βββ utils/ # Utility functions
β βββ __init__.py
β βββ config.py # Configuration management
β βββ text_processing.py # Text processing tools
βββ ReportEngine/ # Multi-round report generation Agent
β βββ agent.py # Agent main logic
β βββ llms/ # LLM interfaces
β βββ nodes/ # Report generation nodes
β β βββ template_selection.py # Template selection node
β β βββ html_generation.py # HTML generation node
β βββ report_template/ # Report template library
β β βββ η€ΎδΌε ¬ε ±ηηΉδΊδ»Άεζ.md
β β βββ εδΈεηθζ ηζ΅.md
β β βββ ... # More templates
β βββ flask_interface.py # Flask API interface
βββ ForumEngine/ # Forum engine simple implementation
β βββ monitor.py # Log monitoring and forum management
β βββ llm_host.py # Forum host LLM module
βββ MindSpider/ # Weibo crawler system
β βββ main.py # Crawler main program
β βββ config.py # Crawler configuration file
β βββ BroadTopicExtraction/ # Topic extraction module
β β βββ database_manager.py # Database manager
β β βββ get_today_news.py # Today's news fetching
β β βββ main.py # Topic extraction main program
β β βββ topic_extractor.py # Topic extractor
β βββ DeepSentimentCrawling/ # Deep sentiment crawling
β β βββ keyword_manager.py # Keyword manager
β β βββ main.py # Deep crawling main program
β β βββ MediaCrawler/ # Media crawler core
β β βββ platform_crawler.py # Platform crawler management
β βββ schema/ # Database schema
β βββ db_manager.py # Database manager
β βββ init_database.py # Database initialization
β βββ mindspider_tables.sql # Database table structure
βββ SentimentAnalysisModel/ # Sentiment analysis model collection
β βββ WeiboSentiment_Finetuned/ # Fine-tuned BERT/GPT-2 models
β βββ WeiboMultilingualSentiment/# Multilingual sentiment analysis (recommended)
β βββ WeiboSentiment_SmallQwen/ # Small parameter Qwen3 fine-tuning
β βββ WeiboSentiment_MachineLearning/ # Traditional machine learning methods
βββ SingleEngineApp/ # Individual Agent Streamlit applications
β βββ query_engine_streamlit_app.py
β βββ media_engine_streamlit_app.py
β βββ insight_engine_streamlit_app.py
βββ templates/ # Flask templates
β βββ index.html # Main interface frontend
βββ static/ # Static resources
βββ logs/ # Runtime log directory
βββ final_reports/ # Final generated HTML report files
βββ utils/ # Common utility functions
β βββ forum_reader.py # Agent forum communication
β βββ retry_helper.py # Network request retry mechanism tool
βββ app.py # Flask main application entry
βββ config.py # Global configuration file
βββ requirements.txt # Python dependency list
```
## π Quick Start (Docker)
### 1. Starting the Project
**Run Command:** Execute the following command to start all services in the **background**:
```bash
docker compose up -d
```
> **Note: Slow image pull speed.** In the original `docker-compose.yml` file, we have provided alternative mirror image addresses as **comments** for you to replace with.
### 2. Configuration Instructions
#### Database Configuration (PostgreSQL)
Configure the database connection information with the following parameters. The system also supports MySQL, so you can adjust the settings as needed:
| Configuration Item | Value to Use | Description |
| :--- | :--- | :--- |
| `DB_HOST` | `db` | Database service name (as defined in `docker-compose.yml`) |
| `DB_PORT` | `5432` | Default PostgreSQL port |
| `DB_USER` | `bettafish` | Database username |
| `DB_PASSWORD` | `bettafish` | Database password |
| `DB_NAME` | `bettafish` | Database name |
| **Others** | **Keep Default** | Please keep other parameters, such as database connection pool settings, at their default values. |
### Large Language Model (LLM) Configuration
All LLM calls use the OpenAI API interface standard. After you finish the database configuration, continue to configure **all LLM-related parameters** so the system can connect to your selected LLM service.
Once you complete and save the configurations above, the system will be ready to run normally.
## π§ Source Code Startup Guide
> If you are new to building Agent systems, you can start with a very simple demo: [Deep Search Agent Demo](https://github.com/666ghj/DeepSearchAgent-Demo)
### System Requirements
- **Operating System**: Windows, Linux, MacOS
- **Python Version**: 3.9+
- **Conda**: Anaconda or Miniconda
- **Database**: PostgreSQL (recommended) or MySQL
- **Memory**: 2GB+ recommended
### 1. Create Environment
#### If Using Conda
```bash
# Create conda environment
conda create -n your_conda_name python=3.11
conda activate your_conda_name
```
#### If Using uv
```bash
# Create uv environment
uv venv --python 3.11 # Create Python 3.11 environment
```
### 2. Install Dependencies
```bash
# Basic dependency installation
pip install -r requirements.txt
# uv version command (faster installation)
uv pip install -r requirements.txt
# If you do not want to use the local sentiment analysis model (which has low computational requirements and defaults to the CPU version), you can comment out the 'Machine Learning' section in this file before executing the command.
```
### 3. Install Playwright Browser Drivers
```bash
# Install browser drivers (for crawler functionality)
playwright install chromium
```
### 4. Configure LLM and Database
Copy the `.env.example` file in the project root directory and rename it to `.env`.
Edit the `.env` file and fill in your API keys (you can also choose your own models and search proxies; see `.env.example` in the project root directory or `config.py` for details):
```yml
# ====================== Database Configuration ======================
# Database host, e.g., localhost or 127.0.0.1
DB_HOST=your_db_host
# Database port number, default is 3306
DB_PORT=3306
# Database username
DB_USER=your_db_user
# Database password
DB_PASSWORD=your_db_password
# Database name
DB_NAME=your_db_name
# Database character set, utf8mb4 is recommended for emoji compatibility
DB_CHARSET=utf8mb4
# Database type: postgresql or mysql
DB_DIALECT=postgresql
# Database initialization is not required, as it will be checked automatically upon executing app.py
# LLM configuration
# You can switch each Engine's LLM provider as long as it follows the OpenAI-compatible request format
# Insight Agent
INSIGHT_ENGINE_API_KEY=
# Insight Agent LLM API BaseUrl, customize API provider
INSIGHT_ENGINE_BASE_URL=
# Insight Agent LLM Model Name, e.g., kimi-k2-0711-preview
INSIGHT_ENGINE_MODEL_NAME=
# Media Agent
...
```
Recommended LLM API Provider: [aihubmix](https://aihubmix.com/?aff=8Ds9)
### 5. Launch System
#### 5.1 Complete System Launch (Recommended)
```bash
# In project root directory, activate conda environment
conda activate your_conda_name
# Start main application
python app.py
```
uv version startup command:
```bash
# In project root directory, activate uv environment
.venv\Scripts\activate
# Start main application
python app.py
```
> Note 1: After a run is terminated, the Streamlit app might not shut down correctly and may still be occupying the port. If this occurs, find the process that is holding the port and kill it.
> Note 2: Data scraping needs to be performed as a separate operation. Please refer to the instructions in section 5.3.
> Note 3: If page display issues occur during remote server deployment, see [PR#45](https://github.com/666ghj/BettaFish/pull/45)
Visit http://localhost:5000 to use the complete system
#### 5.2 Launch Individual Agents
```bash
# Start QueryEngine
streamlit run SingleEngineApp/query_engine_streamlit_app.py --server.port 8503
# Start MediaEngine
streamlit run SingleEngineApp/media_engine_streamlit_app.py --server.port 8502
# Start InsightEngine
streamlit run SingleEngineApp/insight_engine_streamlit_app.py --server.port 8501
```
#### 5.3 Crawler System Standalone Use
This section has detailed configuration documentation: [MindSpider Usage Guide](./MindSpider/README.md)
MindSpider Running Example
```bash
# Enter crawler directory
cd MindSpider
# Project initialization
python main.py --setup
# Run topic extraction (get hot news and keywords)
python main.py --broad-topic
# Run complete crawler workflow
python main.py --complete --date 2024-01-20
# Run topic extraction only
python main.py --broad-topic --date 2024-01-20
# Run deep crawling only
python main.py --deep-sentiment --platforms xhs dy wb
```
## βοΈ Advanced Configuration (Deprecated: Configuration has been unified to the `.env` file in the project root directory, and other sub-agents automatically inherit the root directory configuration)
### Modify Key Parameters
#### Agent Configuration Parameters
Each agent has dedicated configuration files that can be adjusted according to needs:
```python
# QueryEngine/utils/config.py
class Config:
max_reflections = 2 # Reflection rounds
max_search_results = 15 # Maximum search results
max_content_length = 8000 # Maximum content length
# MediaEngine/utils/config.py
class Config:
comprehensive_search_limit = 10 # Comprehensive search limit
web_search_limit = 15 # Web search limit
# InsightEngine/utils/config.py
class Config:
default_search_topic_globally_limit = 200 # Global search limit
default_get_comments_limit = 500 # Comment retrieval limit
max_search_results_for_llm = 50 # Max results for LLM
```
#### Sentiment Analysis Model Configuration
```python
# InsightEngine/tools/sentiment_analyzer.py
SENTIMENT_CONFIG = {
'model_type': 'multilingual', # Options: 'bert', 'multilingual', 'qwen'
'confidence_threshold': 0.8, # Confidence threshold
'batch_size': 32, # Batch size
'max_sequence_length': 512, # Max sequence length
}
```
### Integrate Different LLM Models
The system supports any LLM provider that follows the OpenAI request format. You only need to fill in KEY, BASE_URL, and MODEL_NAME in `config.py`.
> What is the OpenAI request format? Here's a simple example:
>```python
>from openai import OpenAI
>
>client = OpenAI(api_key="your_api_key",
> base_url="https://api.siliconflow.cn/v1")
>
>response = client.chat.completions.create(
> model="Qwen/Qwen2.5-72B-Instruct",
> messages=[
> {
> 'role': 'user',
> 'content': "What new opportunities will reasoning models bring to the market?"
> }
> ],
>)
>
>complete_response = response.choices[0].message.content
>print(complete_response)
>```
### Change Sentiment Analysis Models
The system integrates multiple sentiment analysis methods, selectable based on needs:
#### 1. Multilingual Sentiment Analysis
```bash
cd SentimentAnalysisModel/WeiboMultilingualSentiment
python predict.py --text "This product is amazing!" --lang "en"
```
#### 2. Small Parameter Qwen3 Fine-tuning
```bash
cd SentimentAnalysisModel/WeiboSentiment_SmallQwen
python predict_universal.py --text "This event was very successful"
```
#### 3. BERT-based Fine-tuned Model
```bash
# Use BERT Chinese model
cd SentimentAnalysisModel/WeiboSentiment_Finetuned/BertChinese-Lora
python predict.py --text "This product is really great"
```
#### 4. GPT-2 LoRA Fine-tuned Model
```bash
cd SentimentAnalysisModel/WeiboSentiment_Finetuned/GPT2-Lora
python predict.py --text "I'm not feeling great today"
```
#### 5. Traditional Machine Learning Methods
```bash
cd SentimentAnalysisModel/WeiboSentiment_MachineLearning
python predict.py --model_type "svm" --text "Service attitude needs improvement"
```
### Integrate Custom Business Database
#### 1. Modify Database Connection Configuration
```python
# Add your business database configuration in config.py
BUSINESS_DB_HOST = "your_business_db_host"
BUSINESS_DB_PORT = 3306
BUSINESS_DB_USER = "your_business_user"
BUSINESS_DB_PASSWORD = "your_business_password"
BUSINESS_DB_NAME = "your_business_database"
```
#### 2. Create Custom Data Access Tools
```python
# InsightEngine/tools/custom_db_tool.py
class CustomBusinessDBTool:
"""Custom business database query tool"""
def __init__(self):
self.connection_config = {
'host': config.BUSINESS_DB_HOST,
'port': config.BUSINESS_DB_PORT,
'user': config.BUSINESS_DB_USER,
'password': config.BUSINESS_DB_PASSWORD,
'database': config.BUSINESS_DB_NAME,
}
def search_business_data(self, query: str, table: str):
"""Query business data"""
# Implement your business logic
pass
def get_customer_feedback(self, product_id: str):
"""Get customer feedback data"""
# Implement customer feedback query logic
pass
```
#### 3. Integrate into InsightEngine
```python
# Integrate custom tools in InsightEngine/agent.py
from .tools.custom_db_tool import CustomBusinessDBTool
class DeepSearchAgent:
def __init__(self, config=None):
# ... other initialization code
self.custom_db_tool = CustomBusinessDBTool()
def execute_custom_search(self, query: str):
"""Execute custom business data search"""
return self.custom_db_tool.search_business_data(query, "your_table")
```
### Custom Report Templates
#### 1. Upload in Web Interface
The system supports uploading custom template files (.md or .txt format), selectable when generating reports.
#### 2. Create Template Files
Create new templates in the `ReportEngine/report_template/` directory, and our Agent will automatically select the most appropriate template.
## π€ Contributing Guide
We welcome all forms of contributions!
**Please read the following contribution guidelines:**
- [CONTRIBUTING-EN.md](./CONTRIBUTING-EN.md)
## π¦ Next Development Plan
The system has currently completed only the first two steps of the "three-step approach": requirement input -> detailed analysis. The missing step is prediction, and directly handing this over to LLM lacks persuasiveness.
Currently, after a long period of crawling and collection, we have accumulated massive data on topic popularity trends over time, trending events, and other change patterns across the entire network. We now have the conditions to develop prediction models. Our team will apply our technical reserves in time series models, graph neural networks, multimodal fusion, and other prediction model technologies to achieve truly data-driven public opinion prediction functionality.
## β οΈ Disclaimer
**Important Notice: This project is for educational, academic research, and learning purposes only**
1. **Compliance Statement**:
- All code, tools, and functionalities in this project are intended solely for educational, academic research, and learning purposes
- Commercial use or profit-making activities are strictly prohibited
- Any illegal, non-compliant, or rights-infringing activities are strictly prohibited
2. **Web Scraping Disclaimer**:
- The web scraping functionality in this project is intended only for technical learning and research purposes
- Users must comply with the target websites' robots.txt protocols and terms of use
- Users must comply with relevant laws and regulations and must not engage in malicious scraping or data abuse
- Users are solely responsible for any legal consequences arising from the use of web scraping functionality
3. **Data Usage Disclaimer**:
- The data analysis functionality in this project is intended only for academic research purposes
- Using analysis results for commercial decision-making or profit-making purposes is strictly prohibited
- Users should ensure the legality and compliance of the data being analyzed
4. **Technical Disclaimer**:
- This project is provided "as is" without any express or implied warranties
- The authors are not responsible for any direct or indirect losses caused by the use of this project
- Users should evaluate the applicability and risks of this project independently
5. **Liability Limitation**:
- Users should fully understand relevant laws and regulations before using this project
- Users should ensure their usage complies with local legal and regulatory requirements
- Users are solely responsible for any consequences arising from the illegal use of this project
**Please carefully read and understand the above disclaimer before using this project. Using this project indicates that you have agreed to and accepted all the above terms.**
## π License
This project is licensed under the [GPL-2.0 License](LICENSE). Please see the LICENSE file for details.
## π Support & Contact
### Get Help
FAQ: https://github.com/666ghj/BettaFish/issues/185
- **Project Homepage**: [GitHub Repository](https://github.com/666ghj/BettaFish)
- **Issue Reporting**: [Issues Page](https://github.com/666ghj/BettaFish/issues)
- **Feature Requests**: [Discussions Page](https://github.com/666ghj/BettaFish/discussions)
### Contact Information
- π§ **Email**: hangjiang@bupt.edu.cn
### Business Cooperation
- **Enterprise Custom Development**
- **Big Data Services**
- **Academic Collaboration**
- **Technical Training**
## π₯ Contributors
Thanks to these excellent contributors:
[](https://github.com/666ghj/BettaFish/graphs/contributors)
## π Project Statistics
[](https://github.com/666ghj/BettaFish/stargazers)
[](https://github.com/666ghj/BettaFish/watchers)
[](https://github.com/666ghj/BettaFish/network)
[](https://github.com/666ghj/BettaFish/issues)
[](https://github.com/666ghj/BettaFish/pulls)
[](https://github.com/666ghj/BettaFish/blob/main/LICENSE)
[](https://github.com/666ghj/BettaFish)
[](https://hub.docker.com/)
[English](./README-EN.md) | [δΈζζζ‘£](./README.md)
Say goodbye to traditional data dashboards. In "WeiYu", everything starts with a simple question - you just need to ask your analysis needs like a conversation
302.AI is a pay-as-you-go enterprise AI resource hub that offers the latest and most comprehensive AI models and APIs on the market, along with a variety of ready-to-use online AI applications.
MindSpider Running Example
