BettaFish 微舆开源项目售前分析

1. Project Overview

The official explanation of "micro-public opinion" is: an innovative multi-intelligent public opinion analysis system realized from 0, helping users to break the information cocoon house, restore the original appearance of public opinion, predict the future direction, and assist decision-making. Users only need to put forward analysis requirements like chat, and the system will automatically analyze mainstream social media at home and abroad and a large number of public comments.

One word I understand:

BettaFish is a public opinion analysis model project that combines "social media crawler multi-Agent in-depth research sentiment analysis report generation". its value is not only in the final report, but also in that it shows a set of multi-Agent data analysis architecture that can be disassembled, learned and transformed into a vertical industry analysis system.

2. Official Schematic

2.1 product entrance and system concept

The official stressed in README: bid farewell to the traditional data billboard. In WeChat, everything starts with a simple question. Users put forward analysis requirements like dialogue. The system automatically disassembles, searches, analyzes and generates reports.

2.2 Multi-Agent Framework Diagram

This figure corresponds to the core Agent structure of the system: Query Agent, Media Agent, Insight Agent, Report Agent, and ForumEngine for Agent collaboration.

2.3 MindSpider running example

MindSpider is a BettaFish public opinion crawler system, responsible for topic discovery and multi-platform deep crawling.

2.4 Logo and Extension Project

The project is also associated with the author's newly released MiroFish, which is positioned as "a simple and universal group intelligence engine that predicts everything". The official description refers to BettaFish as "data collection and analysis" and MiroFish as "panoramic prediction", forming a link from raw data to intelligent decision-making.

3. What can it mainly do

3.1 Global Public Opinion Data Collection

BettaFish built-in MindSpider as a public opinion crawler system, using a two-step approach:

1. Search Agent identifies hot news from 13 social media platforms and technical forums such as Weibo, Zhihu, GitHub and Kuan, and maintains a daily topic analysis table.
2. The platform-wide crawler crawls deeply around each topic to obtain fine-grained public opinion feedback.

The platforms supported by the MindSpider Deep Crawling module include:

Pre-sales value:

many enterprise public opinion systems only look at news headlines or platform hot lists. the BettaFish design focus is to drill down to the comment and user feedback layer to help customers see the real public sentiment instead of just looking at media summaries.

3.2 Multi-Agent Division of Labor Analysis

The README defines four types of core agents:

This division of labor is useful for pre-sales: it is not a black box of "LLM shuttles", but rather the splitting of public opinion analysis into stages of collection, search, multimodal understanding, internal data mining, collaborative discussion, and report production.

3.3 Agent Forum Collaboration Mechanism

BettaFish have a more distinctive ForumEngine. The README describes that it will:

-Monitor each Agent speech.

-Introduction of an LLM moderator.

-Let different agents have chained discussions and debates around issues.

-Let Agent read the forum content and adjust the research direction through forum_reader tools.

Presales:

BettaFish multi-Agent does not simply call several models in parallel, but simulates an analysis group: search Agent, media Agent and database Agent first study separately, then the forum host promotes the discussion, and finally the report Agent summarizes. This mechanism helps to reduce the limitations of a single model perspective.

3.4 Multimodal Public Opinion Analysis

The official emphasizes that the system has multi-modal capabilities and can analyze:

-Graphic content.

-Short video content such as trembles and fast hands.

-Weather, calendar, stock and other structured multi-modal information cards in search engines.

Suitable scenarios:

-Brand short video communication analysis.

-Analysis of short video reviews of public events.

-Live/video platform public opinion research.

-Monitoring of new consumer brands with pictures/videos as the main communication form.

3.5 Sentiment Analysis Model Collection

The project contains 'SentimentAnalysisModel/', which integrates multiple sentiment analysis methods:

-BERT Chinese LoRA spinner.

-GPT-2 LoRA fine-tuning.

-Multilingual sentiment analysis.

-Small parameter Qwen3 fine tuning.

-Traditional ML methods.

This means that it does not rely entirely on the big model prompt for sentiment judgment, but has a dedicated sentiment analysis model and alternative configurations.

Pre-sales value:

For public opinion scenarios, emotional orientation is one of the core indicators. BettaFish provide a variety of sentiment analysis implementation, follow-up can be fine-tuned for industry corpus, such as finance, consumer goods, government, cultural tourism, education and so on.

3.6 automatic generation of structured research reports

BettaFish ReportEngine support:

-Template selection.

-Document layout design.

-Space planning.

-Chapter-level JSON generation and validation.

-Document IR intermediate representation.

-HTML rendering.

-PDF export.

-Markdown export.

-Charts to SVG.

-Chart verification and repair.

Official Report Template:

Official sample report:

-Wuhan University Brand Reputation Depth Analysis Report

-HTML sample of pension service development trend report

-PDF example of the report on the development trend of elderly care services

3.7 the integration of private data and public opinion

README mentions that the platform not only analyzes public public opinion, but also provides an interface to support the integration of internal business databases with public opinion data.

Example direction:

-External social media comment on internal customer complaints.

-Industry Heat Enterprise Sales Data.

-Competition sound volume feedback from your own products.

-Public event public opinion customer service work order.

Pre-sales value:

only by looking at public opinion can we know "what is said outside", and only by combining internal data can we know "how much impact this matter has on our business". The BettaFish Insight Agent is suitable for the solution of "internal insight into external trends.

4. A complete analysis process

According to the README, the complete process can be summarized:

Before sales, this process can be described as an "automation research group":

the user raised a public opinion question, and several professional analysts were started inside the system to search, crawl, read databases, watch videos, make emotional judgments, and then cross-discuss in the forum. finally, the rapporteur wrote a formal analysis report.

5. Applicable Scenario

5.1 Brand Reputation and Crisis Public Relations

Suitable for customers:

-Consumer goods brands.

-Educational institutions.

-Wenlu scenic spot.

-Government publicity department.

-Medical, financial, automotive and other highly reputable and sensitive industries.

Analyzable problems:

-Is an event spreading?

-What are the mainstream platforms talking about?

-Are users angry, questioning, supportive or wait-and-see?

-What groups or platforms are the high-risk views coming from?

Do you need a quick response?

-Does the mood improve after the response?

5.2 Market Competition and Competitive Product Monitoring

Suitable for customers:

-New consumer brands.

SaaS companies.

-Internet products.

-Mobile phones, automobiles, home appliances and other public opinion-intensive industries.

Analyzable problems:

-What events did the competition's recent sound volume come from?

-What are the most common points that users complain and praise?

-What is the word-of-mouth gap between us and the competition?

-Which platforms have new trends or potential opportunities?

5.3 Public Hotspots and Policies/Industry Trend Research and Judgment

Suitable for customers:

-Government and public institutions.

-Think tanks/research institutes.

-Media content team.

-Trade associations.

Analyzable problems:

-What is the public feedback after the release of a policy?

-What are the differences in the concerns of different groups?

-Is public opinion changing in the direction of polarization, moderation or shift?

-What are mainstream and reverse narratives?

5.4 Daily Public Opinion Monitoring Report

BettaFish comes with "daily or regular public opinion monitoring report template", suitable for daily/weekly briefing:

-Hot event ranking.

-Platform volume trend.

-Distribution of emotions.

-Summary of key comments.

-Negative risk tips.

-Follow-up tracking recommendations.

5.5 Vertical Industry Data Analytics Engine

README clearly states that "it starts with public opinion, not just public opinion", and for example, it can be transformed into a market analysis system in the financial sector by modifying the API parameters and prompt of the Agent tool set.

Retrofit direction:

6. Not quite the scene

7. Architecture and module disassembly

7.1 Core Module

7.2 Technology Stack

7.3 LLM Configuration Features

Configure APIs for different agents in '.env.example:

-Insight Agent: Kimi K2 is recommended.

-Media Agent: Gemini 2.5 Pro is recommended.

-Query Agent: DeepSeek Chat is recommended.

-Report Agent: Gemini 2.5 Pro is recommended.

-MindSpider Agent: DeepSeek Chat is recommended.

-Forum moderator: Recommended Qwen Plus.

-SQL Keyword Optimizer: Qwen Plus is recommended.

This design shows that the author does not pursue the same model for all tasks, but hopes that different Agents use models suitable for their own tasks.

Pre-sales value:

Customers can configure models hierarchically by cost and capability: cheap models for search/topic extraction, strong models for multimodal and report generation, and stable Chinese models for SQL keyword optimization.

8. How to deploy and use

8.1 Docker Quick Start

The official Docker process:

cp .env.example .env
# 编辑 .env，配置数据库、LLM 和搜索 API
docker compose up -d

Database default parameters:

Post-Startup Access:

http://localhost:5000

8.2 source code startup

Environmental requirements:

-Windows/Linux/macOS.

-Python 3.9.

-Recommended Python 3.11.

-PostgreSQL recommended, MySQL is also supported.

-Memory above 2GB is recommended.

To create an environment:

conda create -n bettafish python=3.11
conda activate bettafish

Installing dependencies:

pip install -r requirements.txt
playwright install chromium

Start the complete system:

python app.py

Access:

http://localhost:5000

8.3 to start the Agent separately

streamlit run SingleEngineApp/query_engine_streamlit_app.py --server.port 8503
streamlit run SingleEngineApp/media_engine_streamlit_app.py --server.port 8502
streamlit run SingleEngineApp/insight_engine_streamlit_app.py --server.port 8501

8.4 MindSpider crawler use

First Cloning Suggested Pull Submodule:

git clone --recurse-submodules https://github.com/666ghj/BettaFish.git
cd BettaFish/MindSpider

If cloned but MediaCrawler empty:

git submodule update --init --recursive

Initialization/Checking:

python main.py --status

Run topic extraction:

python main.py --broad-topic

Run Depth Crawl:

python main.py --deep-sentiment --platforms xhs dy wb

Complete process:

python main.py --complete --date 2024-01-20

It is recommended to use test mode first:

python main.py --complete --test

8.5 Report Retry and Regenerate

If you are not satisfied with the final report, you can just run the Report Engine:

python report_engine_only.py --query "土木工程行业分析"
python report_engine_only.py --skip-pdf
python report_engine_only.py --verbose

You can also use:

-'regenerate_latest_html.py'

-'regenerate_latest_md.py'

-'regenerate_latest_pdf.py'

Re-render the latest chapter JSON or IR.

9. What can I say before sales

9.1 Facing Brand/Market Leader

micro-public opinion is not only to give you a public opinion billboard, but to help you automatically organize a group of AI analysts: some are responsible for the whole network search, some are responsible for short videos and pictures, some are responsible for internal databases, some are responsible for emotional judgment, and finally generate a readable brand reputation or market competition report.

9.2 for PR/Crisis Team

For emergencies, the most important thing is to be quick and accurate: where to start fermenting, which opinions are spreading, whether the user's mood is upgraded, and whether the response strategy is effective. The advantage of Weiyu is that it can drill down into the comment layer and form event research and judgment through multi-Agent cross-analysis.

9.3 for data/AI teams

BettaFish is a multi-agent data analysis project model suitable for learning and secondary development. It does not rely on LangChain/LangGraph and other frameworks, but is implemented from 0 in Python, which is convenient to disassemble Query, Media, Insight, Report, Forum and other modules and replace them with their own models, databases and tools.

9.4 for government and enterprise/industry customers

this kind of architecture can be transformed into an industry public opinion and trend research system: combining public platform data, internal work orders/hotlines/complaints/sales data, allowing AI to automatically form daily/weekly reports and risk tips.

9.5 compliance words that must be added

Current open source repositories are expressly stated to be for learning, academic research, and educational purposes only, and commercial use is strictly prohibited. If the customer wants to use it in a production or commercial project, they need to contact the author for authorization, enterprise customization, or re-evaluation of commercially available alternatives.

10. PoC Recommendations

10.1 PoC Target

PoC is recommended to target:

Verify that the multi-agent public opinion analysis process automates data collection, comment drill-down, sentiment analysis, cross-discussion, and report generation on a given topic, and evaluates compliance, cost, accuracy, and interpretability.

10.2 PoC Scenario 1: Brand Reputation Analysis

Process:

1. Choose a public brand or analog brand that does not use customer sensitive data.
2. Configure the LLM for the Query/Media/Report Agent.
3. Collect limited platform data with test mode.
4. Generate brand reputation report.
5. Manual sampling to verify the source of comments, emotional judgments and conclusions.

Evaluation indicators:

10.3 PoC Scenario 2: Emergency Monitoring

Process:

1. Select historical public events or public cases.
2. Use MindSpider to obtain relevant keywords and platform feedback.

Let ForumEngine drive the multi-agent discussion.

1. Generate public opinion report on crisis public relations.
2. Compare the artificial public opinion briefing.

Evaluation indicators:

-Whether to identify the main points of dispute.

-Whether emotional inflection points are found.

-Whether to distinguish between fact, opinion and speculation.

-Whether reasonable disposal recommendations are given.

-Whether there are hallucinations or over-inferences.

10.4 PoC Scenario 3: Competitive Word-of-Mouth Analysis

Process:

1. Choose 2-3 competing items.
2. Collect user comments and social media discussions respectively.
3. Generate a public opinion report on the market competition pattern.
4. Output strengths/weaknesses/opportunity points.

Evaluation indicators:

-Competition sound volume comparison.

-Clustering of user pain points.

-Emotional tendency contrast.

-Quality of actionable recommendations.

10.5 PoC Preconditions

11. Risks and Precautions

11.1 Commercial Use Risk

This is the most important risk. The README disclaimer clearly states:

-For study, academic research and educational purposes only.

-Any commercial use or for-profit activities is strictly prohibited.

-It is strictly prohibited to be used for illegal, illegal or infringing on the rights and interests of others.

-The use of the results of the analysis for commercial decision-making or profit-making purposes is strictly prohibited.

At the same time the project License is GPL-2.0. The GPL-2.0 itself allows commercial use and redistribution, but the README's additional disclaimer places stronger restrictions on the boundaries of use. Don't simply say "open source so commercial" before sales, you must first make authorization confirmation.

Recommendation:

-For learning and program research: can refer.

-For customer PoC: to qualify for non-commercial research purposes and review terms.

-For commercial delivery: Contact the author for commercial license or custom development.

-For self-research substitution: reference architecture and ideas to avoid direct copying of restricted code.

11.2 crawler compliance risk

The project involves multi-platform crawler, login status and comment crawling, which must be paid attention:

-robots.txt.

-Platform Terms of Service.

-Account wind control.

-Protection of personal information.

-Data minimization.

-Data retention and deletion.

-Whether to allow automated access.

Before sales, we should avoid promising that "any platform can be grasped stably and indefinitely".

11.3 Data Quality Risk

Public opinion analysis relies on data sources. Frequently Asked Questions:

-Platform crawl results in incomplete data.

-Comment on sample bias.

-Popular content is not equal to full public opinion.

-Water Army/Robot/Marketing Number Influence Judgment.

-Multimodal content parsing may be unstable.

-The affective model lacks recognition of satire, metaphor, dialect, and black language.

11.4 LLM Illusion and Report Credibility Risk

The Report Agent generates formal reports, but the LLM may:

-Make up reasons.

-- Exaggerating the trend.

-Confusing facts and speculation.

-Generate inaccurate charts or references.

-Inappropriate advice on sensitive matters.

It is suggested to add manual review and random inspection of evidence chain in PoC.

11.5 Production Engineering Risk

The project is very popular, but not equal to the production level. Additional assessment required:

-Concurrent task scheduling.

-Long mission failure recovery.

-Crawler account pool and proxy pool.

-Logging and monitoring.

-Rights management.

-API throttling.

-Database capacity and indexes.

-Security isolation.

-Report quality stability.

12. Comparison with adjacent schemes

13. Frequently Asked Customer Questions

14. Pre-sales Question List

14.1 Business Issues

-Is the customer monitoring brand, product, competition, policy or emergency?

What platforms need to be covered?

-Focus on sound volume, emotion, risk, transmission path or crowd portrait?

-Is the reporting frequency real-time, daily, weekly, or event-triggered?

-Is the end user PR, marketing, leadership, researcher or customer service team?

14.2 Data and Compliance

-Is it allowed to collect public data of the target platform?

-Does it involve personal or sensitive information?

-Is data desensitization required?

-Do you have a platform account and authorization?

-How long is the data retained?

-Allow access to internal database?

14.3 Technology Environment

-Can I deploy a Python PostgreSQL/MySQL environment?

-Can I use Docker Compose?

-Can I install a Playwright browser?

-Is there an LLM API and a search API available?

-Is there a proxy/network condition to access the target platform?

-Is intranet deployment required?

14.4 Commercial License

-Is it only for study research PoC?

-Is it used for customer commercial projects?

-Do I need to contact the author for authorization?

-Is it acceptable to rewrite based on the schema instead of using the original code directly?

15. My Pre-Sales Judgment

BettaFish is a very suitable for learning and explaining the multi-agent public opinion analysis project. Its strength lies in the integrity of the business closed loop: from topic discovery, social media crawling, comment drilling, sentiment analysis, multi-Agent forum collaboration, to the final HTML/PDF/Markdown report generation, basically covering the main process of the public opinion analysis system.

For pre-sales, the best use of it is not to "take it and sell it", but:

-Use it to explain how multi-Agent can land in real industry scenarios.

-Use it to disassemble the end-to-end architecture of the public opinion analysis system.

-Use it for non-commercial, learning PoC or internal technical verification.

-Use it to inspire customers: how to combine public opinion with internal business data to form automated research reports.

But its boundaries must also be clarified:

-README expressly restricts commercial use.

-High crawler compliance risk.

-Production engineering also requires significant reinforcement.

-The LLM report must be reviewed.

-Multi-platform availability is affected by login, risk control, anti-crawling, and API costs.

One word conclusion:

BettaFish is a good "multi-Agent public opinion analysis reference architecture" and a learning open source project, which is suitable for pre-sales to talk about solutions, do technical verification and initiate industry scenarios. However, if it is to be used for customer commercial delivery, the first step is not to deploy, but to solve the problems of authorization, compliance and production engineering.

16. Quick Sales Summary

17. REFERENCE

-GitHub-666ghj/BettaFish

-BettaFish README

-MindSpider README

-DeepWiki-BettaFish

-Release v3.0.0

-Wuhan University brand reputation depth analysis report example

-HTML Example of the Elderly Care Service Development Trend Report

-PDF Sample of the Report on Trends in Elderly Care Services

-FAQ Summary Issue #700

-MiroFish Prediction Engine