Has anyone tried recursive summarization of documents using LLMs

Question

I'm dissatisfied with how semantic search currently seems to be implemented (statements -> embeddings -> KNN/ANN). Was thinking about attempting to try a different approach involving recursive summarization using a task fine tuned LLM. I wonder if anyone's tried this before, and if so what were the problems with it (aside from the obvious problem of hallucinations). Apologies in advance if this has been discussed before.

spikedoanz · Accepted Answer

as a side note. I'm aware of the approaches using Langchain/CoT to crawl through the document to search for content. This also seems less than ideal due to a lot of processing at inference time. I'm moreso looking for a zero-shot approach like how you can just ask questions to ChatGPT