How to deal with real and recommended chunk size?

The content of some page should become a knowledge for a combination of customGPT with a vector database and/or for AI Assistant (don’t know at the moment, what performs for me better). The goal is a recommendations engine.

After scraping content and embeddings calculation the size of embeddings for single scraped URL is around 20k tokens. What I read in several documentations - recommended chunk size is ~128 tokens, i.e. here.

So I have generally two questions:

  1. What could be a proper way to divide scraped content into such small parts?
  2. With a plan to import content and embeddings into vector database or AI Assistant, how crucial is it for understanding by AI, to use such small chunks?