Qwen: Qwen3 Rerank

Qwen3 Rerank evaluates candidate text against a supplied query and ranking instruction. The exact configured selector is qwen3-rerank. Its role is ordering the supplied candidates for retrieval and document-selection work; the request keeps the query, instruction and candidate passages together so the relevance judgment can be traced back to its inputs.

Prepare documents as an array of strings, with one candidate passage in each entry. Keep a separate mapping from each array position to your own document identifier, title or source location. This lets the application reconnect a ranked result with the original passage without adding undocumented identifier objects to the string array. For a support search, the candidates might be the relevant sections already retrieved from a product guide; the query should state the user's actual information need.

The documented request requires model, documents, query, top_n and instruct. The query and instruct fields are strings, and top_n is an explicitly supplied integer. The published example uses three candidate texts, top_n=2 and an instruction to retrieve relevant passages that answer a web-search query. Use instruct to state the relation that matters for the ranking, while keeping the actual question in query and the material to evaluate in documents.

The operation guide lists a maximum input length of 30K. Keep the candidate passages focused on the text needed to judge their relevance, retaining headings or nearby context when those change the meaning. Preserve source boundaries so a selected passage can still be checked in its original document. The exact request profile and its required fields remain the basis for preparing the input.

The returned structured result contains results entries with index and relevance_score. The index refers to the candidate's position in the original documents array, rather than its position in the ranked result list. Keep the returned ordering and use each index to recover the corresponding input text. Preserve every returned result so the application can inspect the selected candidates and their scores together.

Retain the complete JSON response, including its model, response id and usage.total_tokens. The returned document text can be absent, so keep the original candidate array available when processing the indices. The relevance scores and reported usage have separate roles: scores describe the ranking result, while usage records the quantity reported for the operation. Neither replaces the source passage that supports a later answer.

For a retrieval workflow, use the ranking to select the passages that the next step should inspect or cite. Compare a selected passage with the query and instruction before presenting its contents as evidence. A high position in a result list identifies a candidate for review; the application's eventual answer should remain connected to the actual text and its source identifier.

Review the result with representative queries and the candidate sets used by the application. Check that indices resolve to the intended original texts, that the ordering serves the supplied instruction and that the chosen passages address the question. Keep the exact selector, query, instruction, candidate mapping and complete response together, producing a reusable relevance-selection record for the next retrieval or document step.