One quick tip I totally left out of the ingestion section: watch your API rate limits.
When you first move from the 'Toy' stage to a real database, it's really easy to just loop through 500k chunks and send them to OpenAI or Cohere. You will hit a 429 rate limit error almost immediately. Save yourself the headache and set up a simple queue with exponential backoff before you do your first massive ingestion run