The ai-mreflow from deutschey

YouTubeGPT ft. Matt Wolfe (@mreflow)

AI Chatbot with 100+ videos from YouTuber Matt Wolfe @mreflow

📝 About

💻 How to build

Initial setup
Handle massive data
Embeddings and database backend
Frontend UI with chat
Run app

🚀 Next steps

Deploy
Customizations

🔧 Tools used

👤 Contact

📝 About

Chat with 100+ YouTube videos from any creator in less than 10 minutes. This project combines basic Python scripting, vector embeddings, OpenAI, Pinecone, and Langchain into a modern chat interface, allowing you to quickly reference any content your favorite YouTuber covers. Type in natural language and get returned detailed answers: (1) in the style / tone of your YouTuber, and (2) with the top 2-3 specific videos referenced hyperlinked.

(back to top)

💻 How to build

Note: macOS version, adjust accordingly for Windows / Linux

Initial setup

Clone and install dependencies:

git clone https://github.com/vdutts7/ai-mreflow
cd ai-mreflow
npm i

Copy .env.example and rename to .env in root directory. Fill out API keys:

ASSEMBLY_AI_API_TOKEN=""
OPENAI_API_KEY=""
PINECONE_API_KEY=""
PINECONE_ENVIRONMENT=""
PINECONE_INDEX=""

Get API keys:

AssemblyAI - ~ $3.50 per 100 vids
OpenAI
Pinecone

IMPORTANT: Verify that .gitignore contains .env in it.

Handle massive data

Outline:

Export metadata (.csv) of YouTube videos ⬇️
Download the audio files
Transcribe audio files

Navigate to scripts folder, which will host all of the data from the YouTube videos.

cd scripts

Setup python environemnt:

conda env list
conda activate youtube-chat
pip install -r requirements.txt

Scrape YouTube channel-- replace @mreflow with @ of your choice. Replace <k-last-vids> with the number of videos you want included (the script traverses backwards starting from most recent upload). A new file <your-csv-file>.csv will be created at the directory as referenced below:

python scripts/scrape_vids.py https://www.youtube.com/@<username> `<k-last-vids>` scripts/vid_list/<your-csv-file>.csv

Refer to example.csv inside folder and verify your output matches this format:

Download audio files:

python scripts/download_yt_audios.py scripts/vid_list/<your-csv-file>.csv scripts/audio_files/

We will utilize AssemblyAI's API wrapper class for OpenAI's Whisper API. Their script provides step-by-step directions for a more efficient, faster speech-to-text conversion as Whisper is way too slow and will cost you more. I spent ~ $3.50 to transcribe the 112 videos for Matt Wolfe.

python scripts/transcribe_audios.py scripts/audio_files/ scripts/transcripts

Upsert to Pinecone database:

python scripts/pinecone_helper.py scripts/vid_list/<your-csv-file>.csv scripts/transcripts/

Pinecone index setup I used below. I used P1 since this is optimized for speed. 1536 is OpenAI's standard we're limited to when querying data from the vectorstore:

Embeddings and database backend

Breaking down scripts/pinecone_helper.py :

Chunk size of 1000 characters with 500 character overlap. I found this working for me but obviously experiment and adjust according to your content library's size, complexity, etc.
Metadata: (1) video url and (2) video title

With Pinecone vectorstore loaded, we use Langchain's Conversational Retrieval QA to ask questions, extract relevant metadata from our embeddings, and deliver back to the user in a packaged format as an answer.

The relevant video titles are cited via hyperlinks directly to the video url.

Frontend UI with chat

NextJs styled with Tailwind CSS. src/pages/index.tsx contains base skeleton. src/pages/api/chat-chain.ts is heart of the code where the Langchain connections are outlined.