Multichannel Discovery

Inspiration

The germ of the idea came from looking into the California DMV PDF file and how to link it to the content in the California DMV Youtube channel. At this hackathon, the project started with the intent of discovering the capabilities of Llama Parse, but then morphed into gathering knowledge

What it does

For a given PDF file, perform ask and answer with it, but also get related video links on Youtube.

How we built it

1. Get Llama Parse to run
2. Get Youtube search term summarization from LLM ( either BentoML or OpenAI )
3. Integrate with Youtube
4. Add a Streamlit frontend
5. Integrate with BentoML
6. Integrate with either Vectara or AstraDB. Vectara needed you to preload the document, so was no selected.

Challenges we ran into

1. Instead of GPT-4, I also wanted to use the models available in BentoML, but deployment took an inordinate amount of time.
   
2. I had never used Streamlit in depth until now, so discovering what it could do and working around it workflow was challenging.
3. Wanted to initially use Vectara, but it required you to preload the document, so opted for AstraDB instead.

Accomplishments that we're proud of

Lots of hurdles were overcome and I'm glad it's working.

What we learned

Streamlit was the biggest gain, but knowing about Llama Parse was interesting too.

What's next for Multichaneel Discovery

If there is traction, build out integration to other sources.

Built With

• astradb
• bentoml
• gpt-4
• llama-parse
• llamaindex
• openai
• python
• streamlit
• xxts
• youtube