RAG Financial Platform

A full-stack data platform with Retrieval-Augmented Generation (RAG) workflows, robust security pipelines (PII redaction), and advanced multi-agent orchestration for querying financial data using natural language.

---

🎨 Liquid Glass User Interface

The frontend has been completely revamped into a Bespoke Financial Terminal, featuring a cutting-edge Liquid Glass UI. It embraces a dark, modern aesthetic with deep blurs, high-contrast neon accents, and smooth micro-animations, ensuring zero "AI slop" and providing a premium financial analyst experience.

---

🏗️ Architecture & Core Decisions

Recent architectural shifts have transitioned this platform away from paid API dependencies to a highly optimized, cost-effective hybrid model.

1. Hybrid LLM & Embedding Strategy

• Text Generation (Groq): We leverage the Groq API (using groq/compound) for instantaneous SQL generation and agent reasoning. It acts as a drop-in OpenAI API replacement but executes significantly faster and at zero cost.
• Embeddings (Local Native): Instead of relying on external services (which cause latency and cost) or heavy local servers (like Ollama), we utilize sentence-transformers running directly in the Python process. The vectorizer uses all-MiniLM-L6-v2 (384 dimensions) for lightning-fast, highly accurate semantic search over the database schema.

2. Multi-Agent Pipeline with PII Redaction

We employ a robust orchestrator that sanitizes inputs before routing them to specialized agents.

graph TD
    User([User Query]) --> PII[PII Redactor System]
    PII -- Sanitized Query --> Orchestrator[Agent Orchestrator]
    
    Orchestrator --> |1. Schema & Data Retrieval| RA(Retrieval Agent)
    RA --> |Generates SQL via Groq| DB[(PostgreSQL Data)]
    RA -.-> |FAISS Semantic Search| VS[FAISS Schema Index]
    
    Orchestrator --> |2. Fetch Market Data| EA(Enrichment Agent)
    EA -.-> |yfinance / SEC API| Ext[External APIs]
    
    Orchestrator --> |3. Synthesize Insights| AA(Analysis Agent)
    
    RA --> AA
    EA --> AA
    AA --> |Final Answer & Charts| UI[Liquid Glass Frontend]

Loading

3. PII Reduction System

All inbound user queries pass through the PIIRedactor, a strict regex-based sanitization layer that scrubs sensitive entities (Email, Phone, SSN, Credit Cards, IPv4) before they ever touch an LLM or Vector Store.

---

🗄️ Database Schema

• companies - Company information and metadata
• financial_statements - Income statements, balance sheets, cash flow (5000+ rows)
• portfolio_companies - PE fund portfolio tracking
• performance_metrics - ARR, MRR, churn, CAC, LTV (5000+ rows)
• market_data - Historical stock prices
• query_logs - Query history and debugging

---

🚀 Setup Instructions

Prerequisites

• Python 3.10+ with pip
• Node.js 18+ and npm
• PostgreSQL 13+
• A free API key from Groq Console

1. Database Setup

Create a PostgreSQL database:

createdb rag_data

2. Environment Variables

Copy .env.example to .env and fill in every required value. Never commit .env.

Variable	Required	Description
DATABASE_URL	✅ Yes	Full PostgreSQL connection string (e.g., postgresql://postgres:1234@localhost:5432/rag_data)
GROQ_API_KEY	✅ Yes	Groq API Key (e.g. gsk_...)
OPENAI_API_KEY	✅ Yes	Dummy key (e.g., "groq") for the OpenAI client
OPENAI_API_BASE	✅ Yes	Groq Base URL (https://api.groq.com/openai/v1)
API_KEY	Optional	Secret key for auth. Leave blank for local dev.

3. Backend Setup

cd backend
pip install -r requirements.txt
pip install sentence-transformers  # Required for local embeddings

# Run Alembic migrations to build the tables
alembic upgrade head

# Synthesize mock data (Optional)
python setup_data.py

Start the FastAPI server:

uvicorn api.main:app --reload --host 0.0.0.0 --port 8000

4. Frontend Setup

cd frontend
npm install
npm run dev

Frontend: http://localhost:5173

---

🔒 Security Posture

This project implements the following security controls:

• PII Redaction Engine — All queries are intercepted and sanitized before touching the orchestration layer.
• Anti-Hallucination Guardrails — The SQL generator is strictly prompted to refuse queries that require hallucinated tables/columns, returning a clean error rather than breaking the database.
• SQL Injection Prevention — LLM-generated SQL is parsed, strictly validated to ensure it is a SELECT statement, and row-capped (max 500 rows) via sqlglot.
• Local Embeddings — FAISS and sentence embeddings run entirely locally. Schema data never leaves the server.
• Rate Limiting — API endpoints are protected using slowapi.

---

🛠️ Technical Stack

Backend

• FastAPI — Modern async web framework
• SQLAlchemy 2 & Alembic — ORM and database migrations
• Groq API — Ultra-fast LLM inference (groq/compound)
• Sentence-Transformers — Local embeddings (all-MiniLM-L6-v2)
• FAISS — Vector similarity search
• sqlglot — SQL parsing and validation

Frontend

• React 19 & Vite — High-performance UI
• Vanilla CSS — Custom Liquid Glass / Bespoke Financial Terminal aesthetics

---

📝 License

MIT