Finance RAG - Advanced RAG Pipeline for SEC 10-K Filings

An advanced Retrieval-Augmented Generation (RAG) system that answers questions about Apple and Tesla SEC 10-K annual filings using state-of-the-art NLP techniques.

RAG Architecture

This pipeline implements 6 advanced RAG techniques:

1. Parent-Child Chunking (Dense X) -- Small child chunks (512 chars) for precise retrieval, large parent chunks (2048 chars) for rich context
2. RAG Fusion -- Generates multiple query variations and fuses results with Reciprocal Rank Fusion
3. Stepback Prompting -- Generates abstract queries for broader context retrieval
4. Semantic Routing -- Classifies queries (financial metrics, risk factors, comparisons, etc.) to optimize retrieval strategy
5. Cross-Encoder Reranking -- BGE reranker for precise relevance scoring after initial retrieval
6. Adaptive Generation -- Adjusts response strategy (high/medium/low confidence) based on retrieval quality

Tech Stack

Component	Technology
LLM	Google Gemini 2.0 Flash
Embeddings	BAAI/bge-base-en-v1.5
Reranker	BAAI/bge-reranker-base
Vector DB	ChromaDB

Data Coverage

• Apple 10-K filings: 2018, 2019, 2020, 2021, 2022, 2023, 2024, 2025
• Tesla 10-K filings: 2020, 2021, 2023, 2024, 2025
• 2,016 pages and 1,735 table documents indexed

Example Queries

• "What was Apple's revenue in 2023 compared to 2020?"
• "What are Tesla's main risk factors?"
• "Compare Apple and Tesla's profit margins"
• "How has Apple's R&D spending changed over the years?"