LLM is a fuzzy retriever instead of a reasoning machine

• Many people have the mental impression that LLM such as gpt-4 is capable of advanced reasoning and thinking. Such belief could hinder the progress of whatever goals under which such assumption is held, instead of speeding it up.
• The above statement is very controversial, there are several high quality for and against papers discussing the LLM capabilities.

I am the believer of LLM is a very powerful fuzzy retriever due to it's attention mechanism, whether that amount to reasoning depends on how you define reasoning.

RAG (Retrieval Augmented Generation) is not LLM

• Though LLM can be arguable, if we change the statement to RAG, the statement is largely correct in my opinion.

Many people have the mental impression that LLM with RAG is capable of advanced reasoning and thinking on the materials given. Such belief could hinder the progress of whatever goals under which such assumption is held, instead of speeding it up.

• RAG is clearly not like LLM, where the transformer architecture and hence the attention mechanism is trained on large corpus of data.
• Under most circumstances
  • we do not have enough data nor compute power to train another LLM.
  • fine-tuning does not yield very meaningful result other than certain use case, such as style or tone of the llm.
• The idea that when we put enough documents into a RAG system, magic will happen is plainly wrong.

A better model to think about RAG is a COPY and PASTE model, signal vs noise

• DO NOT just dumb every documents into a RAG system. It will just add noise to the system and confuse the retrieval process.
• Reduce the search space for the RAG system and make each piece of information atomic and useful.

Think of it as:

how do we select the minimal set of knowledge units, such that the RAG system can focus on picking the right piece of information to COPY and PASTE into the LLM as SYSYTEM PROMPT along with the user query?