How many new scientific papers could LLama read, compare and draw conclusions from, analyzing the new information from the papers with the capabilities developed through its training by Llama? I see in the example on the website k=3 for a simple vector store. How many more could it reasonably do to be able to compare across?
Comparisons usually work by extracting summary or answer to a query for each document, and then generating a final answer from those sub-queries.
In this case, if each index was it's own paper, and the max output length of a response is 256 tokens, then it could reasonably compare ~13-14 papers assuming 4096 context size, depending on how long the query string is.
