LucieXLiu

Recently saw a Scale x OpenAI webinar that described some strategies for improving RAG systems. Looks like there are a good number of strategies to try for optimizing the retriever, like optimizing chunk sizes and classifying/ranking chunk relevance, but building a robust RAG system require rigorous experimentation against the specific dataset that feeds the retriever. Not sure how many of these GPTs bothered with this experimentation, and I think you are right that without knowing how that retrieved data is fed into the model they have very few parameters to experiment and optimize with, so they probably just can’t.

Chatbots with retrieval access to scientific literature exist. I tried two of these and built a third one myself. They are good for finding references that might have escaped your attention but getting a coherent, critically-thought-through output from them is difficult. For now the RAG setup can give users the illusion that it is answering with coherent understanding of a body of knowledge, but really it is pretty scatter-brained and the UI should make sure the user goes into the paper to check the quality themselves if they are going to use it as a reference. There is work to be done to improve the retrieval and fine-tuning, but I haven't tried yet. Getting human feedback on these responses also requires expertise in extremely finely-divided subfields and is slightly harder than labeling cats and dogs.

Another issue with the rapid-fire facts Q&A setup is that it is at odds with the scientific method. Science is a dynamic process of observing patterns, proposing hypotheses to explain patterns, and testing the hypotheses with experiments to produce theories. In your question, the "pure fact" probably refers to a theory that is unanimously supported by good experiments, and "BS" is the opposite of that. However, there is a lot of ground between fact and BS because a theory can be supported by early experiments and then proven false by later experiments. The newer the theory is, more likely it is to be sitting somewhere in the middle of this spectrum.

Many experimental fields are having a bit of a reproducibility crisis, and many theoretical fields are divided over theories that are difficult to experiment on. So for the newer literature you really need to judge for yourself, and for the older literature there is real risk that somebody has come up with some new observations that proved it wrong, even if it has gathered a lot of citations over the years.