Query understanding is the set of steps that transform a raw query string into a structured, corrected, expanded representation before retrieval happens.
It typically includes normalisation and tokenisation, spelling correction, stemming or lemmatisation, synonym expansion, entity and attribute extraction, and intent classification. A query like "blak dress undr 3000 sleevless" should reach the retriever as a corrected phrase plus a price filter plus an attribute constraint.
Most perceived "bad relevance" originates here rather than in ranking. If the engine never understood that 3000 was a budget, no amount of re-ranking will produce the right page.