Grounded Financial Document Question Answering (RAG)

A strictly grounded Retrieval-Augmented Generation (RAG) system for answering factual questions over financial documents such as annual reports and regulatory filings.

The system is designed to eliminate hallucinations by enforcing a simple rule:

If an answer is not explicitly present in the documents, it will not be generated.

---

Overview

This project implements an end-to-end RAG pipeline that ingests financial documents, retrieves relevant context using semantic search, and generates answers using a locally hosted large language model.

All responses are grounded in retrieved document text.
If the information required to answer a question is missing, the system refuses to answer.

The entire pipeline runs fully offline, without relying on external LLM APIs.

---

Key Capabilities

• Financial document ingestion (PDF / TXT)
• Semantic chunking with overlap
• Dense vector embeddings and FAISS-based retrieval
• Local LLM inference using Ollama (Mistral)
• Strict grounding and refusal-based hallucination control

---

Technologies Used

• Python 3.10+
• LangChain – RAG orchestration
• FAISS – vector similarity search
• SentenceTransformers – text embeddings
• Ollama – local LLM serving
• Mistral (Ministral-3) – language model

---

System Pipeline

1. Load financial documents from disk
2. Split documents into overlapping text chunks
3. Generate embeddings for each chunk
4. Store embeddings in a FAISS index
5. Retrieve top-k relevant chunks for a query
6. Generate an answer strictly from retrieved context

---

Project Structure

Genai_patham/
├── data/
│   └── docs/                    # Financial source documents
│       ├── nvidia_10k.txt
│       └── NVIDIA-2025-Annual-Report.pdf
│
├── faiss_index/                 # Persisted FAISS vector store
│   ├── index.faiss
│   └── index.pkl
│
├── notebooks/
│   └── training.ipynb           # Development and experimentation
│
├── ingest.py                    # Document ingestion and indexing
├── query.py                     # CLI interface for querying the system
│
├── requirements.txt
├── README.md
└── .gitignore

Setup

1. Clone the repository

git clone https://github.com/Van-der/Financial_RAG_Advisor.git
cd Financial_RAG_Advisor

2. Create and activate a virtual environment

python -m venv .venv
source .venv/bin/activate      # Linux / macOS
# .venv\Scripts\activate       # Windows

3. Install dependencies

pip install -r requirements.txt

4. Install and start Ollama

ollama pull ministral-3
ollama serve

Usage

Step 1: Index documents

Run this once (or whenever documents change):

python ingest.py

This builds and stores the FAISS index in faiss_index/.

Step 2: Query the system

python query.py

Example queries

>> By what percentage did NVIDIA’s revenue grow year-over-year in Fiscal 2025?
>> What major industry shift drove NVIDIA’s performance in Fiscal 2025?
>> What is NVIDIA’s stock price today?

Grounding and Hallucination Control

The system enforces strict grounding:

• Answers are generated only from retrieved document chunks
• No external knowledge or assumptions are allowed
• If the answer is not explicitly present, the system responds with:

Not found in documents.

This behavior is enforced through both prompt constraints and programmatic guardrails.

Current Limitations

• No source citations or page numbers
• CLI-only interface
• Single-domain focus (financial filings)

Possible Extensions

• Add citations (page number + text snippet)
• Streamlit or web-based UI
• Dockerized deployment
• Multi-agent verification using LangGraph
• Support for additional document formats

---

License

This project is released under the MIT License.