I Built a Youtube Chatbot ft. AssemblyAI / Llamaindex / Weaviate

[@matten_zero]

I also participated but I def took the whole week to finish and had a partner to focus on the front end. Good luck to you!

[@sehmiconscious]

CONGRATS on hitting 5K! 💯💥

[@andfanilo]

🎉😁🍰 thank you for the support! Keeps me energized :)

[@CustomGPT]

Very nicely done Fanilo.

[@andfanilo]

Thanks for the support :) hope to keep seeing you around

[@rapunzelle714]

That was super interesting! I got a lot of ideas how I could improve the app I submitted for the hackathon.

[@andfanilo]

Awesome! What did you submit :) ? I'd love to play with it!

[@rapunzelle714]

It's an app to help you practice speaking in a language you are learning. You can select language proficiency, register and grammar corrections. It also tracks custom vocabulary to practice and you can upload texts you want to talk about. Was my first streamlit app I built. I used AssemblyAI to transcribe what the user says, send it to GPT 3.5 for the bot response and then use just gTTS to talk back to the user to keep costs down

[@johngoad]

so cool... you are the best! BTW you can get youtube transcripts for free, so kids... save your allowance. Google will show you how. Transcripts can be Golden Panda Poo when building a focused "expert at that" bot, first find the smartest KZbinr and copy the transcript to a file... then use Langchain to add the doc to your flow. To make the bot even better use prompt templates "You are Fanilo, a funny expert Streamlit developer that is an expert at CSS and explains complicated technical topics in a way that a anyone can understand" Thanks for the videos! I am so sick of looking at my terminal and this helps.

[@andfanilo]

Thanks for sharing all those tips! Definitely would want to try again...but with more than 3h at hand :p

[@sehmiconscious]

Hi - thanks for the video and code. I was able to reuse your Streamlit sandbox and chat code with my own Weaviate index with very few changes (literally just the index name and property fields). I think if you had specified the vectorizer in the schema when you indexed your content (i.e., 'vectorizer': 'text2vec-openai'), then you would have been able to use "with_near_text". (At 6:35 you can see the "vectorizer" is "none".)

[@andfanilo]

👍 eheh glad it worked for you Oh really, did I miss the vectorizer argument in llama-index? That's on me for trying to speedrun this, thanks for sharing! I'll have to give it a try

[@sehmiconscious]

FYI - I set the vectorizer key in the schema at indexing time and the near text queries worked.