Multi-Hop RAG System

📌 What This Project Does

This project builds a Retrieval-Augmented Generation (RAG) system that can answer multi-hop questions. It reads a set of documents, converts them into embeddings, stores them in FAISS, retrieves relevant chunks for a query, and generates final answers using an LLM.

It can answer questions that require combining multiple documents, for example:

• “Which airline had the highest cancellations and what caused them?”
• “Which carrier improved on-time performance despite rising delays?”

---

📌 Steps Performed

1. Load & preprocess documents

• Reads raw text files
• Splits them into clean chunks
• Prepares text for embedding

2. Create embeddings

• Uses OpenAI embedding model to convert text into numeric vectors

3. Store embeddings in FAISS

• Build a FAISS index
• Save metadata (chunk text + source file)

4. Retrieve relevant chunks

• For a given question, retrieve top K relevant text chunks

5. Multi-hop reasoning

• Combine information from multiple sources
• Send to an LLM with a structured prompt
• The LLM performs reasoning using only retrieved data

6. Produce final answer

• LLM outputs a clear answer
• Supports complex multi-step or cross-document questions

---

📌 Technologies Used

• Python
• FAISS (vector search)
• OpenAI Embeddings (text embedding)
• OpenAI GPT Models (answer generation)
• Text preprocessing tools (NLTK / regex)

---

📌 Output

The system outputs:

• Retrieved relevant text chunks
• Final combined answer produced by the LLM
• Uses multi-hop reasoning to connect facts
• No hallucination — answers come only from retrieved context

Example output:

Answer:
Spirit Airlines had the most cancellations in 2024 due to fleet grounding
and operational disruptions. Alaska Airlines and Frontier followed due to
weather-related issues.