fix(reranking): use DashScope native API format

- Use native endpoint /api/v1/services/rerank/text-rerank/text-rerank
- Update request body to { model, input, parameters } format
- Update response parsing for { output, request_id } format
- Add documentation for Chinese users using domestic endpoint

Author: bolechen

.gitignore

@@ -26,3 +26,4 @@ yarn-error.log*
 *.backup
 CLAUDE.md
 AGENTS.md
+test-rerank.ts

CHANGELOG.md

@@ -1,5 +1,15 @@
 # qwen-ai-provider-v5
 
+## 2.0.2
+
+### Patch Changes
+
+- Fix reranking API endpoint and request format
+  - Use DashScope native API endpoint `/api/v1/services/rerank/text-rerank/text-rerank` instead of OpenAI compatible mode
+  - Update request body format to `{ model, input: { query, documents }, parameters: { top_n } }`
+  - Update response parsing for `{ output: { results: [...] }, request_id }` format
+  - Add documentation for Chinese users using domestic endpoint
+
 ## 2.0.1
 
 ### Patch Changes

README.md

@@ -222,6 +222,18 @@ const modelLegacy = qwen.textEmbeddingModel("text-embedding-v3")
 
 You can create models that call the [Qwen reranking API](https://qwenlm.github.io/blog/qwen3-embedding/) using the `.rerankingModel()` factory method. Reranking models improve search relevance by reordering documents based on their semantic similarity to a query.
 
+> **Note for Chinese users (中国用户)**
+>
+> If you're using a Chinese API key from `dashscope.aliyuncs.com`, you need to specify the domestic endpoint:
+>
+> ```ts
+> import { createQwen } from "qwen-ai-provider-v5"
+>
+> const qwen = createQwen({
+>   baseURL: "https://dashscope.aliyuncs.com/compatible-mode/v1",
+> })
+> ```
+
 ```ts
 const model = qwen.rerankingModel("gte-rerank-v2")
 ```
@@ -381,7 +393,7 @@ console.log(JSON.stringify(result.object.recipe, null, 2))
 
 ### generate-obj-reasoning-mdl.ts
 
-```typescript
+````typescript
 import { generateObject, generateText } from "ai"
 ### generate-object-reasoning-mdl.ts
 
@@ -421,7 +433,7 @@ async function main() {
 }
 
 main().catch(console.error)
-```
+````
 
 ### embed-text.ts
 

package.json

@@ -1,7 +1,7 @@
 {
   "name": "qwen-ai-provider-v5",
   "type": "commonjs",
-  "version": "2.0.1",
+  "version": "2.0.2",
   "packageManager": "pnpm@10.2.0",
   "description": "Vercel AI Provider v6 for running Qwen AI models",
   "author": "bolechen <avenger@php.net>",
@@ -64,6 +64,8 @@
     "@types/node": "^22.13.1",
     "@vitest/coverage-v8": "^3.0.5",
     "@vitest/ui": "^3.0.5",
+    "ai": "^6.0.3",
+    "dotenv": "^17.2.3",
     "eslint": "^9.19.0",
     "eslint-plugin-format": "^1.0.1",
     "husky": "^9.1.7",

pnpm-lock.yaml

@@ -16,17 +16,23 @@ const testDocuments = [
 ]
 
 describe("doRerank", () => {
+  let requestUrl: string
   let requestBody: any
   let requestHeaders: Record<string, string>
   let responseBody: any
   let responseHeaders: Record<string, string>
 
   beforeEach(() => {
+    requestUrl = ""
     requestBody = undefined
     requestHeaders = {}
+    // DashScope native API response format
     responseBody = {
-      results: dummyRankingResults,
+      output: {
+        results: dummyRankingResults,
+      },
       usage: { total_tokens: 100 },
+      request_id: "test-request-id-123",
     }
     responseHeaders = {
       "content-type": "application/json",
@@ -35,13 +41,14 @@ describe("doRerank", () => {
 
   function createTestProvider(overrides?: any) {
     return createQwen({
-      baseURL: "https://my.api.com/v1/",
+      baseURL: "https://my.api.com/compatible-mode/v1",
       headers: {
         Authorization: `Bearer test-api-key`,
       },
       ...overrides,
-      fetch: async (_url, init) => {
+      fetch: async (url, init) => {
         // Capture request
+        requestUrl = url as string
         if (init?.body) {
           requestBody = JSON.parse(init.body as string)
         }
@@ -61,6 +68,21 @@ describe("doRerank", () => {
     })
   }
 
+  it("should use DashScope native API endpoint (not compatible mode)", async () => {
+    const provider = createTestProvider()
+    const model = provider.rerankingModel("gte-rerank-v2")
+
+    await model.doRerank({
+      documents: { type: "text", values: testDocuments },
+      query: "talk about rain",
+    })
+
+    // Should use native API endpoint, not compatible mode
+    expect(requestUrl).toBe(
+      "https://my.api.com/api/v1/services/rerank/text-rerank/text-rerank",
+    )
+  })
+
   it("should extract ranking results", async () => {
     const provider = createTestProvider()
     const model = provider.rerankingModel("gte-rerank-v2")
@@ -77,6 +99,18 @@ describe("doRerank", () => {
     ])
   })
 
+  it("should return request_id in response", async () => {
+    const provider = createTestProvider()
+    const model = provider.rerankingModel("gte-rerank-v2")
+
+    const { response } = await model.doRerank({
+      documents: { type: "text", values: testDocuments },
+      query: "talk about rain",
+    })
+
+    expect(response?.id).toBe("test-request-id-123")
+  })
+
   it("should expose the raw response headers", async () => {
     responseHeaders = {
       ...responseHeaders,
@@ -97,7 +131,7 @@ describe("doRerank", () => {
     })
   })
 
-  it("should pass the model, query, and documents", async () => {
+  it("should pass the model, query, and documents in DashScope native format", async () => {
     const provider = createTestProvider()
     const model = provider.rerankingModel("gte-rerank-v2")
 
@@ -108,12 +142,15 @@ describe("doRerank", () => {
 
     expect(requestBody).toStrictEqual({
       model: "gte-rerank-v2",
-      query: "talk about rain",
-      documents: testDocuments,
+      input: {
+        query: "talk about rain",
+        documents: testDocuments,
+      },
+      parameters: {},
     })
   })
 
-  it("should pass topN parameter", async () => {
+  it("should pass topN parameter in parameters object", async () => {
     const provider = createTestProvider()
     const model = provider.rerankingModel("gte-rerank-v2")
 
@@ -125,13 +162,17 @@ describe("doRerank", () => {
 
     expect(requestBody).toStrictEqual({
       model: "gte-rerank-v2",
-      query: "talk about rain",
-      documents: testDocuments,
-      top_n: 2,
+      input: {
+        query: "talk about rain",
+        documents: testDocuments,
+      },
+      parameters: {
+        top_n: 2,
+      },
     })
   })
 
-  it("should pass returnDocuments setting", async () => {
+  it("should pass returnDocuments setting in parameters object", async () => {
     const provider = createTestProvider()
     const model = provider.rerankingModel("gte-rerank-v2", {
       returnDocuments: true,
@@ -144,9 +185,13 @@ describe("doRerank", () => {
 
     expect(requestBody).toStrictEqual({
       model: "gte-rerank-v2",
-      query: "talk about rain",
-      documents: testDocuments,
-      return_documents: true,
+      input: {
+        query: "talk about rain",
+        documents: testDocuments,
+      },
+      parameters: {
+        return_documents: true,
+      },
     })
   })
 
@@ -164,7 +209,7 @@ describe("doRerank", () => {
       query: "talk about rain",
     })
 
-    expect(requestBody.documents).toStrictEqual([
+    expect(requestBody.input.documents).toStrictEqual([
       JSON.stringify(objectDocs[0]),
       JSON.stringify(objectDocs[1]),
     ])

src/__tests__/qwen-reranking-model.test.ts

@@ -193,11 +193,22 @@ export function createQwen(options: QwenProviderSettings = {}): QwenProvider {
     )
 
   // Create a reranking model instance.
+  // Note: Reranking uses DashScope native API, not OpenAI compatible mode.
+  // The baseURL is extracted without /compatible-mode/v1 suffix.
   const createRerankingModel = (
     modelId: QwenRerankingModelId,
     settings: QwenRerankingSettings = {},
-  ) =>
-    new QwenRerankingModel(modelId, settings, getCommonModelConfig("reranking"))
+  ) => {
+    // Extract the base domain from the baseURL (remove /compatible-mode/v1 if present)
+    const rerankBaseURL = (baseURL ?? "https://dashscope-intl.aliyuncs.com/compatible-mode/v1")
+      .replace(/\/compatible-mode\/v1$/, "")
+    return new QwenRerankingModel(modelId, settings, {
+      provider: "qwen.reranking",
+      baseURL: rerankBaseURL,
+      headers: getHeaders,
+      fetch: options.fetch,
+    })
+  }
 
   // Default provider returns a chat model.
   const provider = (modelId: QwenChatModelId, settings?: QwenChatSettings) =>