I'd also strongly recommend Entity Resolution (data matching) to clean up data -- if the data contains real-world entities -- to improve LLM/RAG performance.