In a question-answering (QA) task, "real" information retrieval, rather than document retrieval is required. Effective QA thus involves complicated and time-consuming processing, such as natural language processing and named-entity processing. To reduce the amount of processing needed, the quantity of documents in a database can be narrowed down during an initial stage of the answering procedure. This paper proposes a new evaluation measurement and compares the retrieval accuracy of initial-stage searching that uses "overall" document retrieval and "partial" passage retrieval with the TREC QA data set. The initial search and final result accuracy for various cutoff points defined according to the number of documents or words that are output is evaluated. A variety of experiments demonstrate that middle-length passage-retrieval is effective for QA, and short-length passage-retrieval could improve the accuracy of the final result for a specific question type.