输入一家公司名称，自动生成 8 章节专业研报
MindStack is a multi-agent system that generates institutional-grade financial research reports. Given a company name or ticker, it orchestrates a team of specialized agents to collect financial data, conduct web research, cross-validate findings, and produce a structured 8-section report — all streamed in real time via WebSocket.
Key design choice: Each agent has a single responsibility. The Chief Editor manages the workflow, Researchers fetch data, Writers draft chapters, and Reviewers validate — no single agent does everything, which keeps prompts focused and outputs consistent.
| # | Section | Data Sources |
|---|---|---|
| 1 | Company Overview | Web search + financial APIs |
| 2 | Core Financial Analysis | Xueqiu / FMP financial statements |
| 3 | Valuation Assessment | PE, PB, PS ratios + industry benchmarks |
| 4 | Industry & Competition | Peer comparison via LLM-generated peer lists |
| 5 | Risk Analysis | Web research + financial health indicators |
| 6 | ESG Evaluation | Web search (governance, environmental) |
| 7 | Technical Analysis | Price trends, moving averages |
| 8 | Investment Thesis | Synthesized from all sections above |
┌─────────────────────────────────────────────────────┐
│ Frontend (Next.js) │
│ WebSocket ↔ real-time progress │
└──────────────────────┬──────────────────────────────┘
│ wss://
┌──────────────────────▼──────────────────────────────┐
│ FastAPI Backend (main.py) │
│ ┌────────────────────────────────────────────────┐ │
│ │ ChiefEditorAgent (LangGraph) │ │
│ │ │ │
│ │ browser ──► planner ──► human ──► researcher │ │
│ │ │ │ │
│ │ ┌────────────────────────────┘ │ │
│ │ ▼ │ │
│ │ writer ──► reviewer ──► reviser │ │
│ │ │ │ │ │ │
│ │ └────┬────┘ │ │ │
│ │ ▼ reject │ │ │
│ │ publisher ◄── approve │ │ │
│ └────────────────────────────────────────────────┘ │
└──────────────────────────────────────────────────────┘
| Agent | Role | Key Logic |
|---|---|---|
| ChiefEditorAgent | LangGraph orchestrator | Manages state graph, routes between nodes, handles retry logic |
| BrowserAgent | Web search | Tavily API for multi-source web retrieval |
| PlannerAgent | Report structuring | Generates 8-section outline with sub-questions per section |
| ResearcherAgent | Data collection | Dual-source routing: A-share/HK → Xueqiu API, US → FMP API |
| WriterAgent | Report drafting | Financial-mode prompt with data citation requirements |
| ReviewerAgent | Quality assurance | Cross-validates financial numbers against source data |
| ReviserAgent | Revision | Addresses reviewer feedback, fixes data discrepancies |
| PublisherAgent | Final assembly | Merges chapters, strips duplicates, formats Markdown output |
Query → Ticker Recognition (3-tier)
├── Regex: 6-digit (A-share), 5-digit (HK), letter codes (US)
├── Hardcoded map: 22 Chinese company names → tickers
└── LLM fallback: extract ticker from natural language query
│
▼
Dual-Source Router
├── isdigit() && len ∈ {5,6} → XueqiuDataTool
│ ├── get_stock_overview() → 10+ financial indicators
│ ├── get_financial_statements() → income/balance/cash flow
│ └── get_industry_peers() → peer comparison data
└── else → FinancialDataTool (FMP)
│
▼
Concurrency Control (asyncio.Semaphore)
└── max_parallel=1 (respects LLM API rate limits)
Three-tier fallback ensures robust stock identification:
- Regex patterns — matches
600519,000001,00700,AAPL,TSLA - Hardcoded map — 22 Chinese company names (贵州茅台, 宁德时代, etc.)
- LLM extraction — calls
call_model()for ambiguous natural language queries
- Python 3.10+
- Node.js 18+
- API keys (see below)
# Clone
git clone https://github.com/Xavier7776/multi-agent-financial-research-platform.git
cd multi-agent-financial-research-platform
# Backend
pip install -r requirements.txt
pip install -r multi_agents/requirements.txt
# Frontend
cd frontend/nextjs
npm installcp .env.example .env| Variable | Required | Purpose |
|---|---|---|
OPENAI_API_KEY |
Yes | LLM API key |
OPENAI_BASE_URL |
Yes | LLM endpoint URL |
TAVILY_API_KEY |
Yes | Web search |
JINA_API_KEY |
Yes | Embeddings (Jina AI, free tier available) |
MAX_PARALLEL_RESEARCH |
No | Concurrent research limit (default: 1) |
MAX_SCRAPER_CONTENT_LENGTH |
No | Content truncation threshold (default: 150000) |
# Terminal 1 — Backend
uvicorn main:app --host 0.0.0.0 --port 8000
# Terminal 2 — Frontend
cd frontend/nextjs
npm run devOpen http://localhost:3000, navigate to the research panel, enter a company name, and watch the agents work in real time.
Deployed on Railway with 1GB RAM. Configuration:
# railway.toml
[build]
builder = "NIXPACKS"
[deploy]
startCommand = "uvicorn main:app --host 0.0.0.0 --port $PORT"
[service]
healthcheckPath = "/health"- Embeddings: Jina AI API (zero local memory) — avoids OOM on 512MB-1GB instances
- Concurrency:
asyncio.Semaphore(1)— respects LLM API 2-concurrent rate limit - WebSocket: 20s heartbeat ping to prevent Railway's 30s idle timeout
- Scraper:
MAX_SCRAPER_CONTENT_LENGTH=50000truncation prevents memory blow-up from large PDFs
The frontend communicates with the backend exclusively over WebSocket (/ws).
{
"type": "start",
"task": "贵州茅台 600519 全面分析",
"report_type": "multi_agents",
"tone": "Objective"
}{
"type": "progress",
"agent": "Researcher",
"message": "Fetching financial data for 600519...",
"data": {}
}{
"type": "complete",
"report": "# 贵州茅台 (600519) 深度研究报告\n\n## 1. 公司概览\n..."
}Heartbeat: client sends "ping" every 20 seconds, server responds "pong".
├── backend/ # FastAPI server
│ └── server/
│ ├── app.py # Application entry, routes, WebSocket handler
│ ├── server_utils.py # Config, file ops, multi-agent execution
│ └── websocket_manager.py # Connection management
├── multi_agents/ # Multi-Agent engine
│ ├── agents/
│ │ ├── chief_editor.py # LangGraph state machine
│ │ ├── researcher.py # Data collection + dual-source routing
│ │ ├── writer.py # Section drafting with financial prompts
│ │ ├── reviewer.py # Data accuracy validation
│ │ ├── reviser.py # Feedback-driven revision
│ │ ├── publisher.py # Report assembly + formatting
│ │ ├── editor.py # Task decomposition
│ │ └── utils/ # File I/O, Markdown → PDF/DOCX
│ ├── components/
│ │ ├── xueqiu_finance.py # A-share / HK stock data (pysnowball)
│ │ └── financial_data.py # US stock data (FMP) + ticker extraction
│ ├── main.py # CLI entry for standalone execution
│ └── task.json # Task definition template
├── gpt_researcher/ # Research engine (web scraping, retrieval)
├── frontend/nextjs/ # Next.js frontend
├── outputs/ # Generated reports
├── requirements.txt
├── Procfile
└── railway.toml
A completed report is structured as follows:
outputs/
└── 贵州茅台-600519-20260708/
├── report.md # Full Markdown report
├── report.pdf # PDF export
├── report.docx # Word export
└── financial_data.json # Raw financial data snapshot
| Decision | Rationale |
|---|---|
| LangGraph over custom orchestration | Native retry, state persistence, and visualization |
| Dual-source routing (not unified API) | No single free API covers A-share + US data with sufficient depth |
| LLM ticker fallback (not embedding similarity) | Embedding models lack financial domain training for ticker disambiguation |
| Jina over local embeddings | Eliminates 200MB+ model memory on constrained deployment instances |
Semaphore(1) default |
LLM API typically enforces 2 concurrent requests; single worker avoids 429 errors |
| Financial data JSON export | Enables post-hoc validation and debugging of generated reports |
MIT