Fund Performance Analysis System

An AI-powered fund performance analysis system that enables Limited Partners (LPs) to automatically process fund performance PDF documents and query fund metrics using natural language.

Table of Contents

• Overview
• Tech Stack
• Features
• Setup Instructions
• Environment Variables
• API Testing Examples
• Screenshots
• Known Limitations
• Future Improvements

Overview

This system solves the problem of manually analyzing quarterly fund performance reports. Limited Partners receive PDF documents containing capital calls, distributions, and performance metrics. This system:

1. Automatically processes PDF documents - Extracts tables (capital calls, distributions, adjustments) and text content
2. Stores structured data - Tables go to PostgreSQL, text goes to FAISS vector store
3. Enables natural language queries - Ask questions like "What is the current DPI?" or "Explain Paid-In Capital"
4. Provides accurate answers - Powered by RAG (Retrieval Augmented Generation) with context-aware conversation

Key Capabilities

• Automatic Fund Creation: PDFs are analyzed to create fund entries automatically
• Intelligent Metrics Calculation: DPI, IRR, TVPI, RVPI, MOIC, NAV calculated from transaction data
• Multi-Fund Comparison: Compare performance metrics across different funds
• Context-Aware Chat: Remembers fund context across conversation turns
• Semantic Search: FAISS-powered vector search for relevant document sections

Tech Stack

Backend

• FastAPI - High-performance Python web framework
• PostgreSQL - Relational database for structured data (transactions, funds)
• FAISS - Vector database for semantic search
• Ollama - Local LLM for embeddings and chat responses
• Redis - Caching and session management
• Celery - Background task processing
• SQLAlchemy - ORM for database operations
• pdfplumber - PDF parsing and table extraction

Frontend

• Next.js 14 - React framework with App Router
• TypeScript - Type-safe development
• TailwindCSS - Utility-first CSS framework
• React Query - Data fetching and caching
• Recharts - Data visualization

Infrastructure

• Docker & Docker Compose - Containerization
• Nginx - Reverse proxy and static file serving

Features

1. Document Processing Pipeline

• ✅ PDF upload and automatic processing
• ✅ Table extraction (capital calls, distributions, adjustments)
• ✅ Intelligent table classification using pattern matching
• ✅ Automatic fund creation from PDF metadata
• ✅ Background processing with Celery workers
• ✅ Data validation and cleaning

2. Vector Store & RAG System

• ✅ Text chunking with overlap for context preservation
• ✅ Local embeddings using Ollama (nomic-embed-text)
• ✅ FAISS index for fast semantic search
• ✅ Hybrid search (vector + metadata filtering)
• ✅ Context retrieval with source attribution

3. Query Engine & Chat System

• ✅ Intent classification (definition vs. calculation vs. retrieval)
• ✅ Context-aware conversation history
• ✅ Fund context extraction across conversation turns
• ✅ Metrics included in conversation history
• ✅ Multi-layer search fallback strategy

4. Metrics Calculation

• ✅ DPI (Distributions to Paid-In Capital)
• ✅ IRR (Internal Rate of Return)
• ✅ TVPI (Total Value to Paid-In Capital)
• ✅ RVPI (Residual Value to Paid-In Capital)
• ✅ MOIC (Multiple on Invested Capital)
• ✅ NAV (Net Asset Value)
• ✅ PIC (Paid-In Capital with adjustments)

5. User Interface

• ✅ Fund dashboard with performance charts
• ✅ Multi-fund comparison table
• ✅ Interactive chat interface with conversation history
• ✅ Document search with highlighting
• ✅ Upload interface with progress tracking
• ✅ Responsive design (mobile-friendly)

6. Performance & Optimization

• ✅ Response caching with Redis
• ✅ Gzip compression middleware
• ✅ Rate limiting
• ✅ Database query optimization
• ✅ Pagination for large datasets

Setup Instructions

Prerequisites

• Docker and Docker Compose installed
• At least 4GB RAM available
• Ports 3000, 8000, 5432, 6379, 11434 available

Quick Start (Docker)

1. Clone the repository

   git clone <repository-url>
   cd coding-test-3rd

2. Configure environment variables

   cp .env.example .env
   # Edit .env with your settings (Ollama will be used by default)

3. Start all services

   docker-compose up -d

4. Wait for Ollama to download models (first time only)

   docker-compose logs -f backend
   # Wait for "Ollama models ready" message

5. Initialize the database

   docker-compose exec backend python -m app.db.init_db

6. Access the application

   • Frontend: http://localhost:3000
   • Backend API: http://localhost:8000
   • API Docs: http://localhost:8000/docs

Development Setup (Without Docker)

See SETUP.md for detailed local development instructions.

Environment Variables

Required Variables

Create a .env file in the root directory:

# LLM Configuration (Ollama - Local)
OLLAMA_BASE_URL=http://ollama:11434
OLLAMA_MODEL=llama3.2
OLLAMA_EMBEDDING_MODEL=nomic-embed-text

# Database (configured in docker-compose.yml)
DATABASE_URL=postgresql://funduser:fundpass@postgres:5432/funddb

# Redis (configured in docker-compose.yml)
REDIS_URL=redis://redis:6379/0

Optional Variables

# OpenAI (if you prefer to use OpenAI instead of Ollama)
OPENAI_API_KEY=sk-your-api-key-here
OPENAI_MODEL=gpt-4-turbo-preview
OPENAI_EMBEDDING_MODEL=text-embedding-3-small

# Anthropic (alternative LLM provider)
ANTHROPIC_API_KEY=your-anthropic-key-here

API Testing Examples

1. Upload a Fund Document

curl -X POST "http://localhost:8000/api/documents/upload" \
  -F "file=@files/fund_reports/Velocity_Ventures_Fund.pdf"

Response:

{
  "document_id": 1,
  "filename": "Velocity_Ventures_Fund.pdf",
  "status": "processing",
  "message": "Document uploaded and processing started"
}

2. Search Documents

curl -X POST "http://localhost:8000/api/search/" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "capital calls in 2024",
    "k": 5
  }'

Response:

{
  "results": [
    {
      "content": "Capital Call - Q1 2024: $25,000,000...",
      "metadata": {
        "document_name": "Velocity_Ventures_Fund.pdf",
        "fund_name": "Velocity Ventures Fund",
        "chunk_index": 3
      },
      "score": 0.89
    }
  ]
}

3. Chat Query

curl -X POST "http://localhost:8000/api/chat/query" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "What is the current DPI for Velocity Ventures Fund?",
    "conversation_id": "abc-123"
  }'

Response:

{
  "answer": "The current DPI (Distributions to Paid-In Capital) for Velocity Ventures Fund is **0.40**...",
  "sources": [
    {
      "content": "Total distributions: $40,000,000...",
      "metadata": {...}
    }
  ],
  "metrics": {
    "paid_in_capital": 100000000,
    "distributed_capital": 40000000,
    "dpi": 0.4,
    "irr": 15.2,
    "tvpi": 1.6
  }
}

4. Get Fund Metrics

curl "http://localhost:8000/api/funds/1/metrics"

Response:

{
  "fund_id": 1,
  "fund_name": "Velocity Ventures Fund",
  "metrics": {
    "paid_in_capital": 100000000,
    "distributed_capital": 40000000,
    "dpi": 0.4,
    "irr": 15.2,
    "tvpi": 1.6,
    "rvpi": 1.2,
    "moic": 1.6,
    "nav": 120000000
  }
}

5. Compare Multiple Funds

curl -X POST "http://localhost:8000/api/funds/compare" \
  -H "Content-Type: application/json" \
  -d '{
    "fund_ids": [1, 2, 3]
  }'

Screenshots

1. Chat Interface with Context Awareness

Natural language queries with conversation history and context awareness

2. Multi-Fund Comparison

Compare performance metrics across multiple funds

3. Fund Detail Page with Charts

Detailed fund performance with visual charts and transaction history

4. Document Search

Semantic search across all fund documents

5. Conversation History

Persistent conversation history with automatic title generation

Known Limitations

1. PDF Processing

• Table Detection: Works best with structured tables. Complex or nested tables may require manual review
• Multi-page Tables: Tables spanning multiple pages may be split into separate entries
• Handwritten Content: Cannot process handwritten or image-based PDFs (needs OCR)

2. Metrics Calculation

• IRR Calculation: Requires at least one cash inflow and one outflow. Returns None if insufficient data
• RVPI Calculation: May show 0 for funds that have fully distributed (when NAV = 0)
• Date Parsing: Assumes dates are in consistent formats (MM/DD/YYYY or YYYY-MM-DD)

3. Search & RAG

• Vector Search Accuracy: Depends on embedding model quality (local Ollama vs cloud embeddings)
• Context Window: Limited to last 10 conversation messages to avoid token limits
• Fund Name Matching: Case-insensitive but requires exact name matches (no fuzzy matching)

4. Performance

• First Query: Slower on first run due to model loading (Ollama warm-up)
• Large PDFs: Documents >50 pages may take 30-60 seconds to process
• Concurrent Uploads: Limited to 1 document processing at a time (Celery single worker)

5. UI/UX

• Real-time Updates: Document processing status requires manual refresh
• Mobile Experience: Chat interface works but comparison tables require horizontal scrolling
• Error Messages: Some backend errors show generic messages instead of user-friendly explanations

Future Improvements

1. Enhanced PDF Processing

   • OCR support for image-based PDFs
   • Better multi-page table handling
   • Automatic table header detection
   • Support for more date formats

2. Improved Search

   • Fuzzy fund name matching
   • Multi-fund query support ("Compare DPI for all tech funds")
   • Date range filtering
   • Advanced filters (fund type, vintage year, etc.)

3. UI Enhancements

   • Real-time document processing status (WebSocket)
   • Bulk upload support
   • Export functionality (CSV, Excel)
   • Dark mode

4. Advanced Analytics

   • Fund performance benchmarking
   • Peer group analysis
   • Trend visualization over time
   • Quartile rankings

5. Collaboration Features

   • User authentication and authorization
   • Share conversations and reports
   • Comments and annotations on documents
   • Team workspaces

6. Data Quality

   • Automatic anomaly detection in metrics
   • Data validation rules engine
   • Audit trail for all calculations
   • Manual override capabilities

7. AI Enhancements

   • Fine-tuned models for PE/VC domain
   • Predictive analytics (projected IRR, etc.)
   • Automatic investment memo generation
   • Smart alerts for significant changes

8. Integration & APIs

   • REST API for third-party integrations
   • Webhook support for automation
   • Export to portfolio management systems
   • Integration with DocuSign, Carta, etc.

9. Enterprise Features

   • Multi-tenancy support
   • SSO/SAML authentication
   • Advanced security (encryption at rest)
   • Compliance reporting (SOC2, GDPR)

---

Contributing

Contributions are welcome! Please read ASSIGNMENT.md for the original project requirements.

License

This project is part of a coding challenge. See ASSIGNMENT.md for details.

Support

For issues or questions, please create an issue in the repository.