One phrasing is fragile
A single query embeds to a single point, and that point may sit just outside the cluster holding the answer. The right passage may use different vocabulary than the user did. Multi query retrieval hedges against this by generating several alternate phrasings of the same question and searching with all of them.
How it works
- A language model produces a handful of paraphrases that vary wording and angle.
- Each paraphrase is embedded and retrieved separately.
- The result lists are merged and deduplicated into one candidate set.
Because each phrasing probes a slightly different region of embedding space, the union covers more of the relevant neighborhood than any single query could.
Costs
You pay for one extra generation to make the paraphrases and several searches instead of one. The merged set may also need a reranker to reorder a larger pool of candidates.
Why it matters
Multi query retrieval directly attacks the vocabulary mismatch between how users ask and how documents are written, raising the chance that at least one phrasing lands near the answer.
Key idea
Multi query retrieval generates several phrasings of a question and unites their results, covering more of embedding space to overcome vocabulary mismatch between question and answer.