# What is Multi-Query Retrieval?

Also called query fan-out.

Multi-query retrieval generates several variations of a single user question, runs a search for each, then merges the result sets into one ranked list. The variations differ in phrasing, specificity, or angle, so a document missed by one formulation can still be found by another. It trades extra retrieval cost for higher recall.

The motivation is that any single query is one sample from many valid phrasings, and retrieval is sensitive to that choice. Asking three or five differently worded versions covers more of the space where relevant documents live. Variations are usually produced by a language model, though they can also come from templates such as a keyword-only form, a natural question form, and a form using domain synonyms.

Merging is the part that decides quality. Reciprocal rank fusion is the common default, since it combines ranked lists without needing comparable scores and rewards documents that several queries agreed on. Simple deduplication by identifier keeps the merged list clean, and a reranking pass afterward is normal, because fusion optimizes recall while reranking restores precision at the top.

Costs are real and multiply. Each variation is a separate retrieval, so index load grows with the fan-out factor, latency grows unless the queries run in parallel, and the model call that produces the variations adds its own delay. Fan-out of three to five is typical; beyond that, returns usually diminish while cost keeps rising.

There are also failure modes worth watching. Generated variations can drift off topic and inject irrelevant documents that fusion then promotes because they ranked highly for a bad query. Variations that are near duplicates add cost without adding coverage. Logging the generated queries and periodically checking which ones contributed unique useful results is the practical way to keep the fan-out honest.

## Key points

- Runs several phrasings of one question and merges results
- Reciprocal rank fusion combines lists without shared score scales
- Raises recall at the cost of latency and index load
- Fan-out of three to five is usually the sweet spot
- Off-topic variations can inject irrelevant documents

## In practice

A question about why deployments are slow expands into three queries: one about deployment latency, one about build pipeline duration, and one about release rollout time. The pipeline document ranks first for only the third variation and never appears for the original wording. Fusion places it third overall, and the answer cites it. Total added latency is about four hundred milliseconds, since the three searches run in parallel.

## Related terms

- [Query Expansion](/en/glossary/query-expansion)
- [Query Rewriting](/en/glossary/query-rewriting)
- [Reranking](/en/glossary/reranking)
- [Retrieval Pipeline](/en/glossary/retrieval-pipeline)
- [Recall and Precision](/en/glossary/recall-and-precision)

[Back to the AI Glossary](/en/glossary)