That's a great question. The LLM does not interact with the vector database. The query is more contextualize before going to the LLM. The original query gets augmented/changed but appending the additional context BEFORE it gets to the LLM. Here's a quick example: Original query: "what is my sick leave policy?" RAG query: "using the provided context, answer the question. Question: what is my sick leave policy? Context: <TEXT RETRIEVED FROM DOCUMENTS IN RAG>" Then this RAG query is sent to the LLM. Hope that helps! Let me know if you have any other questions.