Abstract
A closed-loop pipeline improves retrieval of documents for natural-language queries by updating the documents rather than rewriting queries. Production retrieval executions are instrumented with a binary influence signal indicating whether a document influenced an outcome. An offline pressure test runs a bank of queries against a corpus, computes metrics such as Precision@1 and Recall@5, and flags misses where an expected document does not appear within a threshold rank. For each miss, a diagnosis identifies a failure class, including vocabulary mismatch between colloquial phrasing and document terminology. The corpus is enriched by rewriting along one or more axes: adding colloquial triggers to document metadata, rewriting document content to refine guidance when influence is low, or authoring a new document for uncovered query classes. The corpus is re-indexed and the pressure test is re-run to verify extinction of miss classes via an extinction metric.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Closed-Loop Document Enhancement Using Retrieval Failure Signals", Technical Disclosure Commons, (June 30, 2026)
https://www.tdcommons.org/dpubs_series/10654