Quantum Mechanics Vector RAG with Reranking

• This repository contains the implementation of a basic Vector-based Retrieval-Augmented Generation (RAG) pipeline designed to answer questions based on a corpus of Quantum Mechanics (QM) FAQs.
• The key feature of this project is the integration of a cross-encoder reranker to improve the relevance of retrieved documents before they are passed to the final Language Model (LLM).

✍️ Detailed Blog Explaining This Project End-to-End:

• https://www.kaggle.com/code/nanditab35/vectorrag-reranker-for-qm-faq

🚀 Key Features:

• Vector Database: Utilizes a standard vector store (FAISS) to index the high-dimensional embeddings of the QM text corpus.
• Contextual Reranking: Implements a specialized cross-encoder (e.g., jina-reranker-v2) to re-score the top-$K$ documents retrieved from the vector store. This significantly boosts precision by focusing on true contextual relevance rather than just vector proximity.
• Question Answering: Generates grounded, factual answers using the highly relevant, reranked context.
• Domain: Focused on complex, technical Q&A within the field of Quantum Mechanics.

🛠 Architecture Components:

Component	Role	Example Technology/Model
Embedding Model	Converts text into dense vector representations.	all-MiniLM-L6-v2
Vector Store	Stores and performs initial nearest-neighbor search.	FAISS
Retriever	Fetches the top $K$ document candidates based on similarity.	Custom function based on Vector Store API
Reranker	Rescores candidates based on semantic cross-attention score.	jinaai/jina-reranker-v2-base-multilingual
Generator (LLM)	Synthesizes the final answer from the reranked context.	GPT-3.5

💡 Usage and Execution:

• Environment Setup: Install required libraries: !pip install transformers sentence-transformers numpy faiss-cpu
• Data Ingestion: Load the QM text corpus, split it into chunks, and generate embeddings for storage in the Vector Store.
• RAG Query: Execute the pipeline using the query function:
  • User query is embedded.
  • Top $K$ documents are retrieved.
  • Reranker scores the $K$ documents and selects the final top $N$.
  • LLM generates the answer using the top $N$ context pieces.

⚠️ Known Issues:

• Reranker Cost Anomaly: During testing, the jinaai/jina-reranker-v2-base-multilingual configuration was observed to cause disproportionate API call volume to the generator LLM endpoint, necessitating caution and cost profiling if re-enabled.
• LLM generates the answer using the top $N$ context piece

👥 Author

• GitHub: @nanditab35