http://www.semanlink.net/tag/reranker Documents tagged with Reranking http://www.semanlink.net/doc/2024/09/engineering_blog_journey_to_su > TLDR: The decision though to ignore cost and latency led to a very different system than you'll normally see in this kind of work. It takes minutes to answer questions and uses 100k tokens, but that helps us push past human-level on benchmarks > > - A big finding is that the strong effect of chunk size, parsing quality, and embedding choice vanishes with **LLM rerank/context summary** (RCS). It's slower and more compute, but it makes the systems very robust. > - Another finding was how important it is to do multiple retrieval strategies. Using LLMs, you can get query expansion easily by having LLMs rewrite the question multiple ways. We also found exploiting sci lit metadata - i.e., citation graph - helped a lot 2024-09-20T11:18:59Z http://www.semanlink.net/doc/2024/03/benjamin_clavie_sur_x_docume 2024-03-16T10:28:38Z