@wuzhiguocarter/zhipu-ai-provider 0.2.1 → 0.2.3

package/README.md

@@ -74,7 +74,10 @@ import { zhipu } from 'zhipu-ai-provider';
 
 const { text } = await generateText({
   model: zhipu('glm-4.7', {
-    thinking: { type: 'enabled' } // Enable deep thinking
+    thinking: {
+      type: 'enabled',        // Enable deep thinking
+      clear_thinking: true    // Include reasoning process in response
+    }
   }),
   prompt: 'Explain quantum computing in detail',
 });
@@ -90,6 +93,58 @@ const { text: quickText } = await generateText({
 
 - `thinking: { type: 'enabled' }` - Enable dynamic thinking based on task complexity (default for GLM-4.5+)
 - `thinking: { type: 'disabled' }` - Disable thinking for faster, more direct responses
+- `thinking: { clear_thinking: true }` - Include the reasoning process in the response (`reasoning_content` field)
+- `thinking: { clear_thinking: false }` - Hide the reasoning process, only show final answer
+
+### Tool Streaming (GLM-4.6/4.7)
+
+GLM-4.7 and GLM-4.6 models support controlling tool call streaming:
+
+```ts
+const result = await generateText({
+  model: zhipu('glm-4.7', {
+    toolStream: true  // Stream tool call parameters
+  }),
+  tools: {
+    getWeather: {
+      description: 'Get weather information',
+      parameters: z.object({
+        city: z.string(),
+      }),
+      execute: async ({ city }) => `${city} Sunny 25°C`
+    }
+  },
+  prompt: 'What is the weather in Beijing today?'
+});
+```
+
+- `toolStream: true` - Tool call parameters are streamed in chunks for faster feedback
+- `toolStream: false` - Wait for complete tool call before returning
+
+### Response Format
+
+Control the output format of the model:
+
+```ts
+// JSON mode
+const { text } = await generateText({
+  model: zhipu('glm-4-flash'),
+  responseFormat: { type: 'json' },
+  prompt: 'List three fruits in JSON array format'
+});
+
+// Text mode (explicit)
+const { text } = await generateText({
+  model: zhipu('glm-4-flash'),
+  responseFormat: { type: 'text' },
+  prompt: 'Write a poem'
+});
+```
+
+- `{ type: 'text' }` - Plain text output (default)
+- `{ type: 'json' }` - JSON format output (text models only)
+
+**Note:** Vision and reasoning models do not support JSON format.
 
 ## Embedding Example
 ```ts
@@ -124,6 +179,35 @@ const { image, providerMetadata } = await generateImage({
 console.log(providerMetadata.zhipu.images[0].url)
 ```
 
+## Testing
+
+### Unit Tests
+The project includes comprehensive unit tests that use mock data and do not require API calls:
+
+```bash
+# Run all unit tests
+pnpm test src
+
+# Run specific test file
+pnpm test src/zhipu-chat-language-model.test.ts
+```
+
+### Integration Tests
+Integration tests use the real Zhipu AI API and require an API key:
+
+```bash
+# Set up your API key in .env file
+echo "ZHIPU_API_KEY=your-api-key-here" > .env
+
+# Run integration tests
+pnpm test:node
+
+# Run specific integration test
+pnpm test tests/integration/chat/thinking-mode.test.ts
+```
+
+**Note**: Integration tests consume API quota. See [tests/INTEGRATION_TEST_GUIDE.md](tests/INTEGRATION_TEST_GUIDE.md) for detailed information.
+
 ## Features Support
 - [x] Text generation
 - [x] Text embedding
@@ -141,4 +225,4 @@ console.log(providerMetadata.zhipu.images[0].url)
 ## Documentation
 - **[Zhipu documentation](https://bigmodel.cn/dev/welcome)** 
 - **[Vercel AI SDK documentation](https://sdk.vercel.ai/docs/introduction)**
-- **[Zhipu AI Provider Repo](https://github.com/Xiang-CH/zhipu-ai-provider)**
+- **[Zhipu AI Provider Repo](https://github.com/wuzhiguocarter/zhipu-ai-provider)**

package/dist/index.d.mts

@@ -24,7 +24,7 @@ interface ZhipuChatSettings {
      * @example
      * ```ts
      * const model = zhipu('glm-4.7', {
-     *   thinking: { type: 'enabled' }
+     *   thinking: { type: 'enabled', clear_thinking: true }
      * })
      * ```
      *
@@ -32,12 +32,38 @@ interface ZhipuChatSettings {
      * Only supported by GLM-4.5, GLM-4.6, and GLM-4.7 models. When enabled, the model will use
      * chain-of-thought reasoning for complex tasks. Default is controlled by the API (usually "enabled").
      *
-     * - "enabled": Model uses dynamic thinking based on task complexity
-     * - "disabled": Model responds immediately without deep reasoning
+     * - `type: "enabled"`: Model uses dynamic thinking based on task complexity
+     * - `type: "disabled"`: Model responds immediately without deep reasoning
+     * - `clear_thinking: true`: Include reasoning content in the response (reasoning_content field)
+     * - `clear_thinking: false`: Hide reasoning content, only show final answer
      */
     thinking?: {
         type: "enabled" | "disabled";
+        /**
+         * Whether to include the reasoning process in the response.
+         * Only supported by GLM-4.5, GLM-4.6, and GLM-4.7 models.
+         * When true, the response includes a reasoning_content field with the model's thinking process.
+         */
+        clear_thinking?: boolean;
     };
+    /**
+     * Controls whether tool calls use streaming.
+     * Only supported by GLM-4.7 and GLM-4.6 models.
+     *
+     * @example
+     * ```ts
+     * const model = zhipu('glm-4.7', {
+     *   toolStream: true
+     * })
+     * ```
+     *
+     * @remarks
+     * - `true`: Tool calls are streamed in chunks (faster feedback)
+     * - `false`: Tool calls are returned in complete form (wait for full tool call)
+     *
+     * Default behavior is controlled by the API (typically true for streaming requests).
+     */
+    toolStream?: boolean;
 }
 
 type ZhipuEmbeddingModelId = "embedding-2" | "embedding-3" | (string & {});

package/dist/index.d.ts

@@ -24,7 +24,7 @@ interface ZhipuChatSettings {
      * @example
      * ```ts
      * const model = zhipu('glm-4.7', {
-     *   thinking: { type: 'enabled' }
+     *   thinking: { type: 'enabled', clear_thinking: true }
      * })
      * ```
      *
@@ -32,12 +32,38 @@ interface ZhipuChatSettings {
      * Only supported by GLM-4.5, GLM-4.6, and GLM-4.7 models. When enabled, the model will use
      * chain-of-thought reasoning for complex tasks. Default is controlled by the API (usually "enabled").
      *
-     * - "enabled": Model uses dynamic thinking based on task complexity
-     * - "disabled": Model responds immediately without deep reasoning
+     * - `type: "enabled"`: Model uses dynamic thinking based on task complexity
+     * - `type: "disabled"`: Model responds immediately without deep reasoning
+     * - `clear_thinking: true`: Include reasoning content in the response (reasoning_content field)
+     * - `clear_thinking: false`: Hide reasoning content, only show final answer
      */
     thinking?: {
         type: "enabled" | "disabled";
+        /**
+         * Whether to include the reasoning process in the response.
+         * Only supported by GLM-4.5, GLM-4.6, and GLM-4.7 models.
+         * When true, the response includes a reasoning_content field with the model's thinking process.
+         */
+        clear_thinking?: boolean;
     };
+    /**
+     * Controls whether tool calls use streaming.
+     * Only supported by GLM-4.7 and GLM-4.6 models.
+     *
+     * @example
+     * ```ts
+     * const model = zhipu('glm-4.7', {
+     *   toolStream: true
+     * })
+     * ```
+     *
+     * @remarks
+     * - `true`: Tool calls are streamed in chunks (faster feedback)
+     * - `false`: Tool calls are returned in complete form (wait for full tool call)
+     *
+     * Default behavior is controlled by the API (typically true for streaming requests).
+     */
+    toolStream?: boolean;
 }
 
 type ZhipuEmbeddingModelId = "embedding-2" | "embedding-3" | (string & {});

package/dist/index.js

@@ -13986,7 +13986,7 @@ var ZhipuChatLanguageModel = class {
     tools,
     toolChoice
   }) {
-    var _a3;
+    var _a3, _b;
     const warnings = [];
     if (!this.config.isMultiModel && prompt.every(
       (msg) => msg.role === "user" && !msg.content.every((part) => part.type === "text")
@@ -14062,6 +14062,20 @@ var ZhipuChatLanguageModel = class {
         details: "JSON response format schema is only supported with structuredOutputs, provide the schema."
       });
     }
+    if (((_a3 = this.settings.thinking) == null ? void 0 : _a3.clear_thinking) !== void 0 && !this.modelId.match(/^(glm-4\.[567]|glm-4\.5v)$/)) {
+      warnings.push({
+        type: "unsupported-setting",
+        setting: "thinking.clear_thinking",
+        details: "clear_thinking is only supported by GLM-4.5, GLM-4.6, and GLM-4.7 models."
+      });
+    }
+    if (this.settings.toolStream !== void 0 && !this.modelId.match(/^glm-4\.[67]/)) {
+      warnings.push({
+        type: "unsupported-setting",
+        setting: "toolStream",
+        details: "tool_stream is only supported by GLM-4.7 and GLM-4.6 models."
+      });
+    }
     const baseArgs = {
       // model id:
       model: this.modelId,
@@ -14070,19 +14084,25 @@ var ZhipuChatLanguageModel = class {
       do_sample: this.settings.doSample,
       request_id: this.settings.requestId,
       thinking: this.settings.thinking ? {
-        type: this.settings.thinking.type
+        type: this.settings.thinking.type,
+        ...this.settings.thinking.clear_thinking !== void 0 && {
+          clear_thinking: this.settings.thinking.clear_thinking
+        }
       } : void 0,
       // standardized settings:
       max_tokens: maxOutputTokens,
       temperature,
       top_p: topP,
       // response format:
-      response_format: (responseFormat == null ? void 0 : responseFormat.type) === "json" ? { type: "json_object" } : void 0,
+      response_format: responseFormat ? {
+        type: responseFormat.type === "json" ? "json_object" : "text"
+      } : void 0,
       // messages:
       messages: convertToZhipuChatMessages(prompt),
       // tools:
-      tool_choice: toolChoice != null ? toolChoice : "auto",
-      tools: (_a3 = tools == null ? void 0 : tools.filter((tool) => tool.type === "function").map((tool) => {
+      // tool_choice: toolChoice ?? "auto",
+      tool_choice: "auto",
+      tools: (_b = tools == null ? void 0 : tools.filter((tool) => tool.type === "function").map((tool) => {
         var _a4;
         return {
           type: "function",
@@ -14092,8 +14112,10 @@ var ZhipuChatLanguageModel = class {
             parameters: tool.inputSchema
           }
         };
-      })) != null ? _a3 : void 0
+      })) != null ? _b : void 0,
       // TODO: add provider-specific tool (web_search|retrieval)
+      // tool streaming:
+      tool_stream: this.settings.toolStream
     };
     return {
       args: baseArgs,