🎬 YouTube RAG with Gemini + FAISS

An interactive Streamlit app to ask grounded questions about any YouTube video.
Paste a YouTube URL or ID, build a transcript index with Gemini embeddings + FAISS, and query it with Gemini 2.5 Flash.

---

📖 Description

This app implements a classic Retrieval‑Augmented Generation (RAG) pipeline for YouTube:

1. Fetch transcript via youtube-transcript-api (if available).
2. Split into overlapping chunks (LangChain RecursiveCharacterTextSplitter).
3. Embed with Gemini (models/embedding-001) and index in FAISS.
4. Retrieve top‑k chunks (MMR) and answer using Gemini 2.5 Flash, constrained to transcript context.

---

✨ Features

• 🔗 Paste YouTube URL or raw ID; robust ID parser for common formats.
• 🌐 Transcript language selector (en/hi/bn/te/ta/mr).
• ⚙️ Tunables: chunk size, overlap, Top‑K, LLM temperature.
• 🧩 Show supporting chunks for transparency.
• 🧱 Fail‑safes for disabled or missing transcripts.

---

🛠️ Implementation Details

• Embeddings: GoogleGenerativeAIEmbeddings (models/embedding-001)
• Vector store: FAISS from langchain_community
• Retriever: MMR with configurable k
• LLM: ChatGoogleGenerativeAI(model="gemini-2.5-flash")
• Prompting: strict, context‑only answers; says “don’t know” if context is insufficient
• Frontend: Streamlit with wide layout and sidebar controls

Main app file: YT_app.py

---

🚀 Quickstart

1) Install

pip install -r requirements.txt

requirements.txt (minimal):

streamlit
youtube-transcript-api
langchain
langchain-core
langchain-google-genai
langchain-community
faiss-cpu
python-dotenv

2) API key

Create a .env file:

GOOGLE_API_KEY=your_api_key_here

3) Run

streamlit run YT_app.py

Open http://localhost:8501

---

▶️ How to Use

1. Paste a video URL/ID → Build index / Refresh
2. Type your question → Ask
3. Expand Show supporting chunks to see evidence

---

📸 Demos

Overview

URL / ID input + Build index

Ask a question

Sidebar settings

---

🌥 Deploy to Streamlit Cloud (summary)

1. Push repo with YT_app.py, requirements.txt, README.md (and optional runtime.txt: 3.11).
2. In Streamlit Cloud → New app → pick repo → main file YT_app.py.
3. Add Secrets:

   GOOGLE_API_KEY="your_gemini_api_key"

4. Deploy.

Note: FAISS index is ephemeral on Cloud; rebuilding on refresh is expected.

---

🔮 Roadmap

• Cache & persist FAISS indexes
• Multi‑video collections
• Export answers + sources (JSON/Markdown)
• Fallback speech‑to‑text when captions are unavailable

---

📜 License

Souvik Sarkar