Expanding or rewriting a terse or ambiguous query into a form that retrieves better, before generation happens.
LLM rewrites the query using conversational context, then the rewritten query is what actually hits the retriever.
Same LLM used for CAP-105's generation step