Why chat history changes the retrieval problem

Here is the failure mode. You ship a RAG demo. A user asks "what is a retriever?" and the answer is great. Then they ask "and how does it score results?" and the retriever searches for the literal text "how does it score results", which matches nothing useful, because the word "retriever" only exists in the previous turn.

Two reasons. First, the history is full of answer text, which pollutes the query vector with language from previous responses. Second, the history grows unbounded, and after a few turns the retriever is embedding a paragraph instead of a focused question. A standalone question rewrite keeps the query short, specific, and grounded in what the user actually wants next.