YouTube Search

Search YouTube videos and their context.

Motivation

I wanted to learn how to implement RAG (Retrieval-Augmented Generation) without relying on external APIs, using Docker. I also plan to use this tool to speed up my learning using YouTube.

How it works?

The first part is about extracting the content from YouTube. In fact we don't have to download the video in order to search the content. If we are interested only in English language, we can just download the subtitles and process that instead! The following will serach term kubernetes and download subtitles in srt format. However not all videos have en subtitles. We will first download autogenerated subtitles and then download regular subtitles (if available) to replace the autogenerated ones.

search_phrase="kubernetes"
yt-dlp "ytsearch10:${search_phrase}" --write-auto-subs --sub-lang en --convert-subs srt --skip-download
yt-dlp "ytsearch10:${search_phrase}" --write-subs --sub-lang en --convert-subs srt --skip-download

Fine, so we got our subtitles, but what we will do with them? We will convert the .srt files into .txt files, store them in RAG and then use BM25 + FAISS retrievers to find the relevant chunks of the texts.

%% Document QA System Flow
flowchart TD
    A[Input Folder & Question, Document Loading]
    A --> C[Text Splitting RecursiveCharacterTextSplitter]
    C --> D1[BM25 Retriever]
    C --> D2[FAISS Vector Store]
    D1 --> E[Ensemble Retriever BM25 + FAISS ]
    D2 --> E
    E --> F[Top-k Document Retrieval]
    F --> H[Ollama LLM Processing, Answer Generation]

Loading

For creating embeddings on RAG we use sentence-transformers/all-mpnet-base-v2 (HuggingFace) and for answer generation we use llama3.1:8b (via Ollama). The whole pipeline requires something under 8 GB of RAM and it can take aorund 2 mins to execute rag.py locally.

Usage

Local

Navigate to local/. Make sure you have installed ollama on your system, then pull llama3.1:8b and make sure that it is running

curl -fsSL https://ollama.com/install.sh | sh
ollama pull llama3.1:8b
ollama run llama3.1:8b "What is the capital of France?"

After that make sure that your python3.11 is set up and you install all the requirements. Ideally use uv from Astral. Version 3.12 might be problematic.

uv venv --python=3.11
source .venv/bin/activate
uv pip install -r requirements.txt

This will lso install yt-dlp. The following will process the YT subtitles and answer the defined question. Make sure that ffmepg is installed on your computer

question="What is BERT?"
# download subtitles from YT
bash download_yt_subtitles.sh -s "$question" -n 100 -o data/tmp_srt
# convert subtitles to txt
bash convert_srt_to_txt.sh data/tmp_srt data/tmp_txt
# run the RAG and find the answer
python3.12 rag.py \
    --input_folder data/tmp_txt \
    --question "$question" \
    --top_k 10 \
    --max_recursion 100
# cleaning
rm -r data/tmp_*

This will output

Answer: BERT, or Bidirectional Encoder Representations from Transformers, is a popular language model developed by Google. It is a system and tool that understands language better than any other tool in human history, although not as well as humans do. It is freely available for download, allowing anyone to experiment with it and use it to build systems that can solve various problems related to language.
Reference YouTbe videos:
- [What is BERT and how does it work？ ｜ A Quick Review ](https://www.youtube.com/watch?v=6ahxPTLZxU8)
- [BERT Neural Network - EXPLAINED! ](https://www.youtube.com/watch?v=xI0HHN5XKDo)
- [What is BERT？ ｜ Deep Learning Tutorial 46 (Tensorflow, Keras & Python) ](https://www.youtube.com/watch?v=7kLi8u2dJz0)
- [Language Processing with BERT： The 3 Minute Intro (Deep learning for NLP) ](https://www.youtube.com/watch?v=ioGry-89gqE)
- [BERT model in NLP explained ](https://www.youtube.com/watch?v=j2RYX2hvRz0)

Docker

You can also run the whole app using docker-compose

docker-compose up --build 

This will open gradio app on http://localhost:3000 The question can take around 4 mins to process depending on your hardware.

%% Docker Compose Service Architecture
graph TD
    Redis["Redis
    Image: redis:alpine
    Ports: 6379:6379
    Healthcheck: redis-cli ping every 5s"]
    
    Volume_ollama([Volume: ollama])
    
    Ollama["Ollama
    Image: ollama/ollama
    Ports: 11434:11434
    Runs llama3.1:8b model"]
    
    Backend["Backend
    Build: ./backend
    Ports: 8000:8000
    Env: API_USER, API_PASS, OLLAMA_HOST"]
    
    Gradio["Gradio App
    Build: ./gradio_app
    Ports: 3000:7860"]
    
    Backend -->|"depends_on (service_healthy)"| Redis
    Gradio -->|depends_on| Backend
    Ollama -->|mounts| Volume_ollama
    Backend -->|"OLLAMA_HOST=ollama:11434"| Ollama
    Backend --> Redis

    classDef service stroke:#333,stroke-width:2px;
    classDef volume stroke:#333,stroke-width:2px;
    class Redis,Ollama,Backend,Gradio service;
    class Volume_ollama volume;

Loading

Contibuting

Contributions are welcome! Please open an issue or submit a pull request.