Here is an example: I have 300 embedding on a topic (lets say Chemistry). I want to create a summary comprising 10 top points in those 300 embeddings. How do I do that?
I have one solution in mind (below) but looking for better solutions.
I create 1-2 sentence summary of all 300 embeddings and pass those embeddings at one time as context with a prompt asking to create a 10 point summary. The assumption is that 1-2 lines summary of 300 embeddings would be in token limit (4000 tokens?).
Any other thoughts?
If the summary of any individual embedding doesn’t change on a case by case basis, you could keep a one or two sentence summary of each each embedding with the full text you are using for the search
Then when you find the top 10 records based on the semantic search of the big text field, you can output your pre-saved summaries from the small text field
You could get GPT to summarize each embedding as a one off exercise and save them locally - or you could get a human to write them in advance.
This will only work if the summary for any given embedding doesn’t need to change based on the question.
If a single embedding is a large blocks of text that can answer lots of different questions (and the summary changes), then I think your idea might be the best option
This would only work if the summary (of a given embedding) doesn’t change