Query rewriting: improve the search question without inventing its answer

An intent-preserving rewrite can expand an abbreviation, resolve a reference from authorized conversation context, or translate a question into a corpus's vocabulary. It should not quietly replace an uncertain premise with an invented fact. Keep the original query as a baseline or parallel candidate where appropriate. Log the original, rewrite, rewriter configuration, and retrieval outcome under the application's data policy. A successful fluent rewrite is not evidence that the resolved entity or time range was correct.

Decomposition creates subquestions whose answers may need to be combined. Some are independent and can run concurrently; others depend on a first result, such as identifying a project before searching its owner. The original self-ask work studies such compositional reasoning, while Rewrite-Retrieve-Read studies query adaptation. These are specific methods rather than proof that every question benefits from more queries. Retain the relationship between a subquestion, its supporting sources, and the final claim so a wrong intermediate assumption can be diagnosed.

Bound the number of rewrites, subquestions, retrieval calls, retries, and candidate rows. At a stable request rate, backend call demand grows with the mean executed calls per request even when calls overlap in wall-clock time. Parallel searches reduce only the independent portion of latency and still consume concurrency and bandwidth. Dependent hops require additional serial stages. Timeouts and partial results need an explicit policy: synthesize from sufficient remaining evidence, retry within a budget, ask for clarification, or abstain.

A union can contain many duplicate passages, different versions of one document, and query-specific score scales. Do not compare raw similarity scores from unrelated query formulations as though they were calibrated probabilities. Deduplicate by the intended evidence identity, rerank under a declared policy, and preserve which queries produced each candidate. Every derived search must retain the same principal, permission scope, corpus snapshot, and freshness requirements unless the application explicitly authorizes another scope. A fallback search is not an excuse to drop these constraints.

Create tests with ambiguous names, negation, temporal constraints, compound comparisons, and questions that need only one precise search. Compare original-only, rewrite-only, and bounded multi-query variants with a fixed corpus and generator. Measure authorized evidence recall after deduplication and packing, task accuracy, unsupported claims, and cost. A rewrite can raise retrieval recall while lowering answer quality by introducing a mistaken premise or excessive distractors. Inspect failures by stage rather than crediting every final improvement to the rewriter.

The queue lab isolates fanout pressure as abstract work demand. Reset, set λ=2 and μ=4, predict q(2)=0, then use four Steps. Separately Reset, set λ=4 and μ=3, predict q(2)=6, then use four Steps. Both start with four work units. It has no queries, parallel search graph, or document scores, and its separate M/M/1 residence time has additional stochastic assumptions. Use it to see why more offered work can outgrow service, then measure the actual per-request fanout and concurrency limits.