Return arguments from function calling with OpenAI API when streaming?

Question

I've made a simple OpenAI API example with function calling. I'm only using function calling to format the response, I'm not calling multiple functions or any external APIs.

When I don't stream the response I can return the function arguments, which is the data that I need.

In my NextJS route handler:

export async function POST(request: Request) {
  try {
    const openai = new OpenAI({
      apiKey: process.env["OPENAI_API_KEY"],
    });
    const response = await openai.chat.completions.create({
      model: "gpt-4",
      // stream: true,
      messages: [
        {
          role: "user",
          content: "Give me 5 questions and answers for a pub quiz",
        },
      ],
      tools: [
        {
          type: "function",
          function: {
            name: "get_questions_and_answers",
            description: "Get questions and answers",
            parameters: simpleJsonSchema,
          },
        },
      ],
      tool_choice: {
        type: "function",
        function: { name: "get_questions_and_answers" },
      },
    });
    return Response.json(
       JSON.parse(
         response.choices[0].message.tool_calls?.[0].function.arguments || "",
       ),
    );
  } catch (serverError) {
    console.error({ serverError });
    throw new Error();
  }
}

simpleJsonSchema.json:

{
  "type": "object",
  "properties": {
    "getQuestions": {
      "type": "array",
      "items": {
        "type": "object",
        "properties": {
          "Question": {"type": "string"},
          "Answer": {"type": "string"}
        },
        "required": ["Question", "Answer"]
      }
    }
  },
  "required": ["getQuestions"]
}

Response from API:

{"getQuestions":[{"Question":"What is the capital of Australia?","Answer":"Canberra"},{"Question":"Who wrote 'To Kill a Mockingbird'?","Answer":"Harper Lee"},{"Question":"What is the highest peak in the world?","Answer":"Mount Everest"},{"Question":"Who is known as the 'Father of Computers'?","Answer":"Charles Babbage"},{"Question":"What is the largest ocean in the world?","Answer":"Pacific Ocean"}]}

This is fine when developing locally, however when deployed to Vercel the request sometimes times out. I've tried to add streaming as this is the recommended solution:

const response = await openai.chat.completions.create({
  model: "gpt-4",
  stream: true,
  messages: [
    {
      role: "user",
      content: "Give me 5 questions and answers for a pub quiz",
    },
  ],
  tools: [
    {
      type: "function",
      function: {
        name: "get_questions_and_answers",
        description: "Get questions and answers",
        parameters: simpleJsonSchema,
      },
    },
  ],
  tool_choice: {
    type: "function",
    function: { name: "get_questions_and_answers" },
  },
});

const stream = OpenAIStream(response);
return new StreamingTextResponse(stream);

However now the response has a lot of unnecessary data. And when I try to JSON.parse on the client I get errors.

Response from API:

{"tool_calls":[ {"id": "call_IhxvzkZ5EsmZpHc6tOznTmzb", "type": "function", "function": {"name": "get_questions_and_answers", "arguments": "{\n  \"getQuestions\": [\n    {\n      \"Question\": \"Question 1\",\n      \"Answer\": \"Answer 1\"\n    },\n    {\n      \"Question\": \"Question 2\",\n      \"Answer\": \"Answer 2\"\n    },\n    {\n      \"Question\": \"Question 3\",\n      \"Answer\": \"Answer 3\"\n    },\n    {\n      \"Question\": \"Question 4\",\n      \"Answer\": \"Answer 4\"\n    },\n    {\n      \"Question\": \"Question 5\",\n      \"Answer\": \"Answer 5\"\n    }\n  ]\n}"}}

As far as I can see the docs only cover using useChat but I have some particular requirements so I need to handle the fetching and form state myself: https://sdk.vercel.ai/docs/api-reference/use-chat

Why am I getting invalid JSON?

Here is a repository which reproduces the issue: https://github.com/jameschetwood/openai-function-calling

Answer 1

this is the response you are getting:

{"tool_calls":[ {"id": "call_HRxqlP3yzeHsoN43tMyZjMlr", "type": "function", "function": {"name": "get_questions_and_answers", "arguments": "{\n  \"getQuestions\": [\n    {\n      \"Question\": \"What is the capital city of France?\",\n      \"Answer\": \"Paris\"\n    },\n    {\n      \"Question\": \"Who painted the Mona Lisa?\",\n      \"Answer\": \"Leonardo da Vinci\"\n    },\n    {\n      \"Question\": \"What is the largest planet in our solar system?\",\n      \"Answer\": \"Jupiter\"\n    },\n    {\n      \"Question\": \"What is the national flower of England?\",\n      \"Answer\": \"Rose\"\n    },\n    {\n      \"Question\": \"Which country is famous for its tulips?\",\n      \"Answer\": \"Netherlands\"\n    }\n  ]\n}"}}

I used https://jsoneditoronline.org/ to auto correct the json and it just adds "]}". for some reason openai is not sending correct json response. you have to add it

accumulatedText += "]}";

then response works:

this is too specific error. if openai updates its response api, it might send the json data correctly. so a better approach would be parsing in try/catch

try {
      const parsed = JSON.parse(accumulatedText);
      console.log({ parsed });
    } catch (error) {
      // you should error for each specific case
      accumulatedText += "]}";
      console.log("correct accumulatedText in catch block", accumulatedText);
    }