FinRAG : Financial Document Intelligence

A production-grade Retrieval-Augmented Generation (RAG) system for financial document analysis.
Built with LangChain · FAISS · Sentence Transformers · Mistral-7B · Streamlit

---

Architecture Overview

PDF Upload
    │
    ▼
FinancialPDFLoader          ← PyMuPDF + pdfplumber (tables!)
    │
    ▼
FinancialChunker            ← RecursiveCharacterTextSplitter (800 chars, 150 overlap)
    │
    ▼
FinancialVectorStore        ← sentence-transformers/all-MiniLM-L6-v2 → FAISS index
    │
    ▼
Semantic Retrieval          ← Top-K similarity search (cosine distance)
    │
    ▼
FinancialLLM                ← Mistral-7B (HF Inference API) or GPT-3.5
    │
    ▼
Grounded Answer + Citations ← Page numbers, source chunks, keyword highlights

---

Quickstart

1. Clone & Setup

git clone <your-repo>
cd financial-rag

# Create virtual environment
python -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

2. Get a Free API Key

HuggingFace (Free):

1. Go to https://huggingface.co/settings/tokens
2. Create a token with read access
3. Paste it in the app sidebar

OpenAI (Paid):

• Get key from https://platform.openai.com/api-keys

3. Run

streamlit run app.py

Open: http://localhost:8501

---

Project Structure

financial-rag/
├── app.py                  # Streamlit UI (main entry point)
├── rag_pipeline.py     # Core RAG pipeline
├── requirements.txt
└── README.md

---

Features

Feature	Details
PDF Loading	PyMuPDF + pdfplumber with table extraction
Chunking	Recursive character splitting (800 chars, 150 overlap)
Embeddings	sentence-transformers/all-MiniLM-L6-v2 (384-dim)
Vector DB	FAISS (flat L2 index, cosine normalized)
LLM	Mistral-7B-Instruct via HuggingFace Inference API (free)
Fallback LLM	GPT-3.5-turbo via OpenAI
Financial NLP	60+ keyword detection across 5 categories
Citations	Page numbers + chunk scores in every answer
Summarization	One-click structured financial summary

---

Sample Questions to Try

• "What are the key risks mentioned in this report?"
• "Summarize the revenue trends with specific numbers"
• "What is the company's debt situation?"
• "What guidance or outlook does management provide?"
• "Are there any red flags or concerns raised?"
• "What business segments does the company operate in?"

---

Configuration

Parameter	Default	Description
LLM Provider	huggingface	huggingface or openai
Top-K Chunks	5	Number of retrieved chunks per query
Chunk Size	800	Characters per chunk
Chunk Overlap	150	Overlap between consecutive chunks
Embedding Model	all-MiniLM-L6-v2	Fast, accurate 384-dim embeddings

---

Tech Stack

• LangChain — RAG orchestration, text splitting, document loading
• FAISS — Facebook AI Similarity Search (vector database)
• sentence-transformers — State-of-the-art text embeddings
• Mistral-7B — Open-source LLM via HuggingFace Inference API
• Streamlit — Production-ready Python web UI
• PyMuPDF / pdfplumber — PDF parsing with table support

---

Multi-document support

# Call ingest() multiple times — FAISS merges indexes
pipeline.ingest("report_2023.pdf")
pipeline.ingest("report_2024.pdf")

Persist the vector index

pipeline.vector_store.save("./faiss_index")
pipeline.vector_store.load("./faiss_index")

Custom financial keywords

# In rag_pipeline.py, add to FINANCIAL_KEYWORDS:
FINANCIAL_KEYWORDS["esg"] = ["carbon", "sustainability", "ESG", "net zero"]

---

This project directly demonstrates:

• Generative AI : LLM-powered financial Q&A
• Information Retrieval : Semantic search over financial documents
• Information Extraction : Keyword detection, citation extraction
• Financial Domain : Built for annual reports, earnings, 10-K filings
• Production Engineering : Modular pipeline, error handling, configurable parameters