This repo uses openai embeddings and the pinecone vector database to generate and query embeddings from a VTT file.
The purpose of this repo was to implement semantic search to provide an extra resource in understanding Andrej Karpathy's latest video: Let's build GPT, however it is general enough to use for any transcript.
shoutout to miguel's yt-whisper library for helping with the youtube
transcription. The data/ in this repo was generated using the small model.
pip install -r requirements.txt
cp .env.sample .env
you'll need an API keys from openai & pinecone
OPENAI_KEY=***
PINECONE_KEY=***
PINECONE_ENVIRONMENT=*** (Optional)
Head over to pinecone and create an index with dimension 1536
the data in this repo was generated from Let's build GPT using yt-whisper
/data/karpathy.vtt- contains the raw VTT file/data/karpathy_embeddings.csv- contains the dataframe with the embeddings. you can use this file to directly seed your pinecone index
this will save an embedding csv file as {file_name}_embeddings.csv
python embed_vtt.py generate --vtt-file="data/karpathy.vtt"
python embed_vtt.py upload --csv-embedding-file="data/karpathy_embeddings.csv"
python embed_vtt.py query --text="the usefulness of trill tensors"
sample output
0.81: But let me talk through it. It uses softmax. So trill here is this matrix, lower triangular ones. 00:54:52.240-00:55:01.440
0.81: but torches this function called trill, which is short for a triangular, something like that. 00:48:48.960-00:48:55.920
0.80: which is a very thin wrapper around basically a tensor of shape vocab size by vocab size. 00:23:17.920-00:23:23.280
0.79: I'm creating this trill variable. Trill is not a parameter of the module. So in sort of pytorch 01:19:36.880-01:19:42.160
0.79: does that. And I'm going to start to use the PyTorch library, and specifically the Torch.tensor 00:12:54.320-00:12:59.200
This script is open-source and licensed under the MIT License.