Using LLMs to evaluate other LLMs sounds like a it would be dumb, but LLMs work in mysterious ways. I’ve found this approach useful though. In the context of RA

Using LLMs to evaluate other LLMs sounds like a it would be dumb, but LLMs work in mysterious ways. I’ve found this approach useful though. In the context of RAG, using an LLM to evaluate whether a context chunk is relevant to answer a question is a nice complement to using the vector embedding semantic similarity search. Sometimes prompting the LLM gives better results than vector similarity.