RAG can handle all sorts of different questions, it's really appropriate for extracting information from a document. An interesting question is "when the question requires using more many distant elements from doc, is RAG still relevant?"
@aymericroucher
-
RAG retrieval failures due to random variations, need more tests
By
–
Well spotted!
Both variations (blue and green ) are also due to random variations. For instance the RAG fails to retrieve relevant snippets on 1 single example. The tests I ran were not numerous enough yet to smooth these variations out. But maybe I'll run more. -
RAG system reduces token input from 128k to 2k
By
–
Exactly!
To complement on the end: thanks to the RAG system, the model was fed around 2k tokens each time, down from the 128k tokens of the original document. -
RAG beats GPT-4-Turbo in long document retrieval test
By
–
Information Retrieval: Who wins, GPT-4-Turbo or a RAG based on GPT4? I extended the "Needle in a Haystack" test created by @GregKamradt & the result is clear: 𝗥𝗔𝗚 𝘄𝗶𝗻𝘀 > its edge becomes clear for the longest document sizes. @huggingface Links below