RAG System for 10-K Document Q&A

A question-answering system for Apple's 2024 and Tesla's 2023 10-K filings. Built for the RAG + LLM assignment.

You'll need a free HuggingFace token from https://huggingface.co/settings/tokens

Local setup:

cd rag_project
pip install -r requirements.txt

# Add your HuggingFace token
cp .env.example .env
# Edit .env: HF_TOKEN=your_token_here

python run.py

This creates outputs/answers.json with all 13 answers.

How It Works

1. Parse PDFs and split into chunks (800 characters each)
2. Create embeddings using sentence-transformers
3. Store in FAISS vector database
4. For each question:
   • Find top 15 similar chunks
   • Re-rank to get best 5
   • Send to LLM with context
   • Get answer with citations

Models used:

• Embeddings: all-MiniLM-L6-v2
• Re-ranker: ms-marco-MiniLM-L-6-v2
• LLM: Qwen 1.5B via HuggingFace API

The Required Function

The assignment asks for answer_question() that returns answers with sources:

from src.rag_pipeline import answer_question

result = answer_question("What was Apple's revenue in 2024?")
# {'answer': '$391,036 million', 'sources': ['Apple 10-K', 'Item 8', 'p. 28']}

For questions it can't answer (like stock predictions), it returns empty sources.

Files

• src/ - Main code (document loading, embeddings, retrieval, LLM interface)
• run.py - Runs all 13 evaluation questions
• rag_colab.ipynb - Notebook version for Colab/Kaggle
• requirements.txt - Python packages needed

What It Does (Assignment Requirements)

• Parses both PDFs keeping track of doc name, section, and page number
• Uses open-source embeddings (no paid APIs)
• Stores in FAISS vector database
• Retrieves top 5 chunks using two-stage search
• Uses open-access LLM (not GPT-4 or Claude)
• Cites sources like ["Apple 10-K", "Item 8", "p. 28"]
• Refuses to answer out-of-scope questions
• Works in Colab/Kaggle notebooks