Deeper explanation

[JpEncausse]

Hello,
I like the way you did it ! This project is 1 year old, is it still relevant or did you change your usage ? Because AI is moving so fast.

I was wondering if I correctly understand your process :

1. Create a column for embeddings
2. Convert some fields of a row into an embedding. Store it into the new column

The user ask a question :

1. Ask an LLM to create a "fake row"
2. Create an embeddings for this row
3. Compare distance between this embeddign and the table
4. Build an answer based on top X matching rows.

I didn't understand the 3. wan you explain where/how this cosine similarity works in the code ?
If I have 3000 rows do you iterate on all to run this algorithm ? I don't find how it works. Is it efficient ?

Thanks !