spartan-minhbui / rag-for-video

Question Answering Chatbot with Video

Question Answering Chatbot with Video

Overview

• This is a simple Chatbot that can help user to QA about information of a video.
• For the current version:
  • chatbot can only receive Youtube video url as input.
  • just allow user to input message as text, not image yet
• Chatbot using GPT4o as base LLM.
• Demo: demo

Process flow:

• This chatbot basically use idea of RAG, the pipeline has the following steps:
  • First: It will ask user to input a YouTube video url
    • Video Processor:
      • After receive the url, it will download it to the temporal directory using tempfile lib of python.
      • Then it will convert the video into images (it's also base on the video fps of your setting) and convert video audio into text.
    • Retriever Processor:
      • This processor will receive the folder path of image frames and audio, and it will index the data.
  • Second: After processing the video (it might take you some minutes to finish), user now can input text message and ask about the information in the video.

Installation

• Install required packages: pip install -r requirements.txt
• Install ffmpeg: conda install conda-forge::imageio-ffmpeg

Running

• For the current version, it is highly recommended that you should running with Terminal, because gradio app may still have some bugs.

Terminal

• You can test the bot in terminal by running: python launch.py
• You also have to set the OPENAI_API_KEY env variable to use gpt4o

Gradio

• Also, you can run the chatbot with UI by running: python gradio_app.py

Proposal

• For current version, the chatbot does not handle images as input message, but in I think we can implement this feature by:
  • Step 1: With each input message, use gpt4o to generate the summarize for that image.
  • Step 2: Use that summarized text to and retrieve a list of relevant images using retriever processor.
  • Step 3: Use these retrieved images + input text of user and input it to gpt4o, and return to user the response.