Inspiration

We love RAG applications (using GenAI to search documents) but what we think is cooler is doing RAG on videos.

Watching a video, summarizing it, and then grounding answer to exact moments on the video is the goal of this project.

What it does

You load a video, it will summarize it and enable you to ask questions using Snowflake Arctic model. With the answer a citation will come with the exact moment the answer was mentioned. Automatic playback!.

How we built it

Using Snowflake Arctic model, Replicate and Streamlit.

Challenges we ran into

Chunking video data is complex, specially taking care of the timestamps for playback capabilities.

Accomplishments that we're proud of

The video chunking function, and the playback citations are the best accomplishments of this project.

What we learned

Arctic model is a great start, but the context window is too small!

What's next for Arctic - Video Citations

Adding more analysis on the videos, like topic chunking, fact sub-chunking. And topic/fact classification and video comparisons.

For example: Grab 2 videos ,ask a question and compare what each video says about the same thing.

Built With

  • arctic
  • llamaindex
  • streamlit
Share this project:

Updates