Paper: chain of abstraction (Gao 2024)

[SamHollings]

https://arxiv.org/abs/2401.17464

RAG, is arguably a subset of "Tool-Augmented LLM" - in our case the "tool" is the vector database, which the LLM uses to get a better answer. In this paper they show how you can get better performance out of mathematical problem solving by first abstracting the problem, prior to passing it to the tool. See the image below.

It's possible we could use some of the thinking here in RAG, e.g. have an LLM process the submitted question, break it up, reformat it, etc. (see also RAG fusion).