https://github.com/shivamarora1/youtube-gpt

Youtube GPT
https://github.com/shivamarora1/youtube-gpt

llm mistral-7b nltk qdrant rag retrieval-augmented-generation streamlit summarization youtube

Last synced: 23 days ago
JSON representation

Youtube GPT

• Host: GitHub
• URL: https://github.com/shivamarora1/youtube-gpt
• Owner: shivamarora1
• License: mit
• Created: 2024-04-21T12:02:27.000Z (10 months ago)
• Default Branch: main
• Last Pushed: 2024-05-18T09:49:12.000Z (9 months ago)
• Last Synced: 2024-11-18T13:28:58.786Z (3 months ago)
• Topics: llm, mistral-7b, nltk, qdrant, rag, retrieval-augmented-generation, streamlit, summarization, youtube
• Language: Python
• Homepage: https://youtube-ai-gpt.streamlit.app/
• Size: 226 KB
• Stars: 2
• Watchers: 2
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# Youtube GPT: Converse with Youtube video

![resized_speed_up](https://github.com/shivamarora1/youtube-gpt/assets/28146775/0d5fb99d-b3f6-4e2e-a8bb-d12e515ec71f)

This [Streamlit](https://streamlit.io/) application helps you in summarizing YouTube videos, making it easier to digest content efficiently. Additionally, it allows users to pose follow-up questions related to the video, and the application generates pertinent responses, enhancing engagement and understanding.

### Steps to run on local:

---

1. Create and activate virtual environment

```

python3 -m venv .venv

source .venv

```

2. Download all dependencies

```

pip install -r requirements.txt

```

3. Run application

```

streamlit run main.py 

```

### App overview

---

1. Enter Youtube video link in left side bar.

2. Using Youtube Transcription api application will fetch transcriptions of Youtube video.

3. Fetched transcriptions are sent to `Mistral-7B-instruction` model for summarization.

4. Embeddings of transcriptions are generated using `all-MiniLM-L6-v2` model

5. Generated embeddings are stored in `Qdrant` in memory vector database.

6. When you follow question relevant similar Youtube video context is fetched from vector database and that context is sent to `Mistral-7B-instruction` model along with question.

7. `Mistral-7B-instruction` uses fetched background context and gives answer of asked question. 





    #### Limitations

    1. Youtube videos larger that 30 minutes length are not supported.

    2. Youtube video only in english language are only supported.

### Architecture diagram

![arch](https://github.com/shivamarora1/youtube-gpt/assets/28146775/5bfd6cf4-34b9-4558-8a75-f4360ba5e16d)