Financial Document Analyzer

A multi-agent system for analyzing financial documents (PDFs) using CrewAI and GPT-4. Upload a quarterly report or financial statement, and get back detailed analysis including verification, metrics extraction, investment recommendations, and risk assessment.

Screenshots

Web UI

Analysis Results

Swagger API Documentation

What it does

• Reads PDF financial documents (10-K, quarterly reports, etc.)
• Runs 4 AI agents that each specialize in different analysis
• Stores results in PostgreSQL
• Has a simple web UI + REST API

Setup

1. Clone and add your API keys to .env:

cp .env.example .env
# edit .env with your OPENAI_API_KEY and SERPER_API_KEY

2. Run with Docker:

docker-compose up --build

3. Open http://localhost:8000 and upload a PDF

That's it. The database tables get created automatically on first run.

How to use

Web UI

Go to http://localhost:8000, upload PDF, optionally add a specific question, click analyze.

API

curl -X POST http://localhost:8000/analyze \
  -F "file=@quarterly_report.pdf" \
  -F "query=What are the key financial risks?"

Response includes:

• verification - document authenticity check
• financial_analysis - key metrics and trends
• investment_recommendations - buy/hold/sell thesis
• risk_assessment - identified risks and mitigations

Check API health

curl http://localhost:8000/health

Architecture

4 CrewAI agents working sequentially:

1. Verifier - checks document is legit and complete
2. Financial Analyst - extracts metrics, analyzes trends
3. Investment Advisor - gives investment recommendation
4. Risk Assessor - identifies financial/operational risks

Results get saved to PostgreSQL (analysis and analysis_results tables).

Background jobs go through Celery + Redis.

Stack

• Python 3.11
• FastAPI
• CrewAI 0.130.0
• OpenAI GPT-4 Turbo
• PostgreSQL 15
• Redis (for Celery)
• Docker

Files

main.py          - FastAPI app, API endpoints
agents.py        - CrewAI agent definitions  
task.py          - Task definitions for each agent
tools.py         - PDF reader tool
db_models.py     - SQLAlchemy models
celery_app.py    - Celery configuration
worker.py        - Celery worker
index.html       - Web UI

Bugs I fixed

The original code had several issues that I identified and fixed:

1. Import errors - CrewAI 0.130+ changed import paths (from crewai import Agent not from crewai.agents)
2. Tool pattern wrong - Had to use BaseTool from crewai.tools with _run() method
3. Agents couldn't find files - Fixed by pre-reading PDF content server-side and injecting into prompts
4. SQLAlchemy issues - JSONB import path changed, metadata is reserved keyword (renamed to result_metadata)
5. Missing validation - Added PDF header checks, file size limits, content-type verification
6. No database persistence - Added PostgreSQL storage for analysis results
7. Generic error handling - Added proper HTTP status codes (400, 413, 415, 500)
8. Circular LLM assignment - Original had llm=llm, fixed with proper ChatOpenAI initialization
9. Wrong parameter name - tool= changed to tools= (list of tools)
10. max_iter=1 - Agents gave up after 1 try, increased to reasonable values
11. Missing agent goal parameter - Verifier and Risk Assessor had syntax errors
12. PDFReader import missing - Added from llama_index.readers.file import PDFReader
13. Unprofessional backstories - Rewrote agent backstories to be more professional
14. Incomplete task descriptions - Rewrote task expected outputs with clear structure

Notes

• PDF content is truncated to ~15KB before sending to agents (keeps costs down)
• Each analysis takes 30-60 seconds depending on document size
• Flower dashboard at http://localhost:5555 for monitoring Celery tasks

Environment variables

See .env.example for all options. Main ones:

• OPENAI_API_KEY - your OpenAI key
• SERPER_API_KEY - for web search (optional)
• DATABASE_URL - PostgreSQL connection string