deepseek-reasoner : APICallError with no tools params in request

[cht527]

DEPDENCIES

"dependencies": {
    "@ai-sdk/deepseek": "^0.2.8",
    "@ai-sdk/openai": "^1.3.9",
    "@ai-sdk/react": "^1.2.8",
    "ai": "4.2.10"

MODEL CONFIGURE

export enum LLM_TYPE {
  QWEN_PLUS = 'qwen-plus',
  DEEPSEEK_R1 = 'deepseek-reasoner',
}
export const customMiddleware: Record<LLM_TYPE, LanguageModelV1Middleware> = {
  [LLM_TYPE.DEEPSEEK_R1]: extractReasoningMiddleware({ tagName: 'think' }),
  [LLM_TYPE.QWEN_PLUS] : {}
};

export const customModel: Record<LLM_TYPE, LanguageModelV1> = {
  [LLM_TYPE.DEEPSEEK_R1]: createDeepSeek({apiKey: DEFINE_ENV.DEEPSEEK_API_KEY,baseURL:DEFINE_ENV.DEEPSEEK_BASE_URL })(LLM_TYPE.DEEPSEEK_R1),
  [LLM_TYPE.QWEN_PLUS]: createOpenAI({ apiKey: DEFINE_ENV.QIANWEN_DASHSCOPE_API_KEY,  baseURL: DEFINE_ENV.QIANWEN_DASHSCOPE_API_URL})(LLM_TYPE.QWEN_PLUS)
};

export const getLLM = (modelId = LLM_TYPE.QWEN_PLUS) =>  wrapLanguageModel({
  model:  customModel[modelId],
  middleware: customMiddleware[modelId],
})

LOGIC in streamText function

maxSteps: modelId!==LLM_TYPE.DEEPSEEK_R1 ? 2 : 1,
tools:  modelId!==LLM_TYPE.DEEPSEEK_R1  ? {
          getInformation: tool({
            description: `get information from your knowledge base to answer questions.`,
            parameters: z.object({
              question: z.string().describe('the users question'),
            }),
            execute: getInformationFromKnowledgeBase,
          })}
   : undefined,
onFinish: async ({ response }) => {
            if (session.user && session.user.id) {
              try {
                
                const assistantId = getTrailingMessageId({
                  messages: response.messages.filter(
                    (message) => message.role === 'assistant',
                  ),
                });

                if (!assistantId) {
                  throw new Error('No assistant message found!');
                }

                const [, assistantMessage] = appendResponseMessages({
                  messages: [userMessage],
                  responseMessages: response.messages,
                });

                await saveMessages({
                  messages: [
                    {
                      id: assistantId,
                      chatId: id,
                      role: assistantMessage.role,
                      // content's datastruct same with parts
                      content: assistantMessage.parts,
                      createdAt: new Date(),
                    },
                  ]
                })
    
              } catch (error) {
                console.log(error);
                
                console.error("Failed to save chat");
              }
            }
        },
        
        experimental_telemetry: {
          isEnabled: true,
          functionId: "stream-text",
        },
      });

      result.consumeStream();

      result.mergeIntoDataStream(dataStream, {
        sendReasoning: true,
      })

STEPS

1. pnpm dev, run project in localhost:3000
2. select chat model （qwen-plus in my project），which will use tools call in the progress，and the result is ok
3. select reasoning model （deepseek-reasoner in my project， which will cause error，the detail as shown blow。

ERROR LOG

ERROR: createDataStreamResponse APICallError [AI_APICallError]: deepseek-reasoner does not support Function Calling.
    at eval (webpack-internal:///(rsc)/./node_modules/.pnpm/@[email protected][email protected]/node_modules/@ai-sdk/provider-utils/dist/index.mjs:708:14)
    at process.processTicksAndRejections (node:internal/process/task_queues:95:5)
    at async postToApi (webpack-internal:///(rsc)/./node_modules/.pnpm/@[email protected][email protected]/node_modules/@ai-sdk/provider-utils/dist/index.mjs:608:28)
    at async OpenAICompatibleChatLanguageModel.doStream (webpack-internal:///(rsc)/./node_modules/.pnpm/@[email protected][email protected]/node_modules/@ai-sdk/openai-compatible/dist/index.mjs:491:50)
    at async wrapStream (webpack-internal:///(rsc)/./node_modules/.pnpm/[email protected][email protected][email protected]/node_modules/ai/dist/index.mjs:6337:35)
    at async fn (webpack-internal:///(rsc)/./node_modules/.pnpm/[email protected][email protected][email protected]/node_modules/ai/dist/index.mjs:5523:25)
    at async eval (webpack-internal:///(rsc)/./node_modules/.pnpm/[email protected][email protected][email protected]/node_modules/ai/dist/index.mjs:544:22)
    at async _retryWithExponentialBackoff (webpack-internal:///(rsc)/./node_modules/.pnpm/[email protected][email protected][email protected]/node_modules/ai/dist/index.mjs:345:12)
    at async streamStep (webpack-internal:///(rsc)/./node_modules/.pnpm/[email protected][email protected][email protected]/node_modules/ai/dist/index.mjs:5478:15)
    at async fn (webpack-internal:///(rsc)/./node_modules/.pnpm/[email protected][email protected][email protected]/node_modules/ai/dist/index.mjs:5848:9)
    at async eval (webpack-internal:///(rsc)/./node_modules/.pnpm/[email protected][email protected][email protected]/node_modules/ai/dist/index.mjs:544:22) {
  cause: undefined,
  url: 'https://api.deepseek.com/chat/completions',
  requestBodyValues: {
    model: 'deepseek-reasoner',
    user: undefined,
    max_tokens: undefined,
    temperature: 0,
    top_p: undefined,
    frequency_penalty: undefined,
    presence_penalty: undefined,
    response_format: undefined,
    stop: undefined,
    seed: undefined,
    messages: [
      [Object], [Object],
      [Object], [Object],
      [Object], [Object],
      [Object], [Object],
      [Object], [Object],
      [Object], [Object],
      [Object], [Object],
      [Object], [Object],
      [Object]
    ],
    tools: undefined,
    tool_choice: undefined,
    stream: true
  },
  statusCode: 400,
  responseHeaders: {
    'access-control-allow-credentials': 'true',
    'cf-cache-status': 'DYNAMIC',
    'cf-ray': '92e8167699c5f080-DFW',
    connection: 'keep-alive',
    'content-length': '151',
    'content-type': 'application/json',
    date: 'Fri, 11 Apr 2025 05:29:42 GMT',
    server: 'cloudflare',
    'set-cookie': 'COOKIE VALUE; path=/; expires=Fri, 11-Apr-25 05:59:42 GMT; domain=.deepseek.com; HttpOnly; Secure; SameSite=None',
    'strict-transport-security': 'max-age=31536000; includeSubDomains; preload',
    vary: 'origin, access-control-request-method, access-control-request-headers',
    'x-content-type-options': 'nosniff',
    'x-ds-trace-id': 'dfcbe5aadf4e8be6c55c590e50b20c60'
  },
  responseBody: '{"error":{"message":"deepseek-reasoner does not support Function Calling.","type":"invalid_request_error","param":null,"code":"invalid_request_error"}}',
  isRetryable: false,
  data: {
    error: {
      message: 'deepseek-reasoner does not support Function Calling.',
      type: 'invalid_request_error',
      param: null,
      code: 'invalid_request_error'
    }
  },
  [Symbol(vercel.ai.error)]: true,
  [Symbol(vercel.ai.error.AI_APICallError)]: true
}

message params in ERROR LOG

Since the tools params would be undefined when the model is deepseek-reasoner(LLM_TYPE.DEEPSEEK_R1), I doubt that the messages causes this error, I mark the suspicious message in picture. because I compared the same operation in https://chat.vercel.ai/ . There has no message contains type: "step-start"

[0xrinegade]

@uwularpy

[ailunc]

Hi! @cht527
Here are several approaches you can try:
1. Disable Function Calling for Deepseek Completely:
Since deepseek-reasoner does not support function calling, ensure that no function calling-related parameters are being sent when it is selected. This may involve:
• Bypassing Middleware:
Either conditionally disable the extractReasoningMiddleware when using DEEPSEEK_R1 or adapt it so that it does not insert any extra instructions (such as a "step-start" marker) that imply function calls.
• Modifying Data Stream Merge:
Avoid merging the data stream with sendReasoning: true for deepseek. For example, you could conditionally pass the parameter like:

result.mergeIntoDataStream(dataStream, {
  sendReasoning: modelId !== LLM_TYPE.DEEPSEEK_R1,
});

This way, no reasoning-related function call data is appended when using deepseek.

[cht527]

Hi! @cht527 Here are several approaches you can try: 1. Disable Function Calling for Deepseek Completely: Since deepseek-reasoner does not support function calling, ensure that no function calling-related parameters are being sent when it is selected. This may involve: • Bypassing Middleware: Either conditionally disable the extractReasoningMiddleware when using DEEPSEEK_R1 or adapt it so that it does not insert any extra instructions (such as a "step-start" marker) that imply function calls. • Modifying Data Stream Merge: Avoid merging the data stream with sendReasoning: true for deepseek. For example, you could conditionally pass the parameter like:

result.mergeIntoDataStream(dataStream, {
  sendReasoning: modelId !== LLM_TYPE.DEEPSEEK_R1,
});

This way, no reasoning-related function call data is appended when using deepseek.

@ailunc Hi，This is not the valid way to solve the problem, it simply stops receiving reasoning content. The 'step-start' already happens before I switch the model to DEEPSEEK_R1