Mini RAG (Retrieval-Augmented Generation) App

Welcome! This is a hands-on interview task to assess your ability to build and debug a basic Retrieval-Augmented Generation (RAG) system using Python, FAISS, and language model embeddings.

🧠 Objective

You are given a minimal RAG system that allows users to query a small set of technical documents. The system uses a FastAPI backend with a FAISS index and sentence-transformer-based embeddings.

🔧 However, the current system is returning irrelevant or inaccurate results. Your task is to identify and fix the retrieval logic to improve the quality of answers.

---

📂 Project Structure

rag_app_debugging/
│
├── app.py              # FastAPI app with /ask endpoint
├── rag_utils.py        # FAISS + embedding logic (this is your main focus)
├── client.py           # Sends test query to the app
├── data/
│   └── docs.txt        # Text corpus (technical content)
├── requirements.txt    # Python dependencies
└── README.md           # This file

---

🛠️ Setup Instructions

1. Clone the project and create a virtual environment

   python3 -m venv .venv
   source .venv/bin/activate  # or `.venv\Scripts\activate` on Windows

2. Install dependencies

   pip install -r requirements.txt

3. Start the FastAPI server

   uvicorn app:app --reload

4. Run the client to send a query

   python client.py

---

🎯 Your Task

The current system does a basic embedding + FAISS L2 search. You need to improve the context retrieval quality using any of the following techniques (feel free to implement one or more):

Suggested Improvements

• 🔹 Improve document chunking (e.g., use sentence/paragraph-based or sliding window)
• 🔹 Replace IndexFlatL2 with IndexFlatIP (cosine similarity)
• 🔹 Normalize embeddings before indexing/search
• 🔹 Return top-k matches (k > 1) and optionally re-rank them
• 🔹 Use a better embedding model (e.g., all-MiniLM-L12-v2, intfloat/e5-base-v2)
• 🔹 Add cross-encoder re-ranking if time permits

---

✅ Success Criteria

• The /ask endpoint returns relevant and meaningful results for typical technical queries.
• You explain the changes you made and why they improve the results.
• Your code is clean, modular, and well-commented.

---

🧪 Example Queries

Try asking the system:

• "What is the name of the war operation by Israel?"
• "What is the reason for the war?"
• "What is President Trump's stance on the war?"
• "What could end the war?"

---

🤝 Good Luck!

We’re not expecting a production-grade solution — just a thoughtful and focused approach to improving retrieval quality with clear reasoning. Feel free to leave comments in the code or discuss trade-offs.

If you have any questions during the session, just ask.