Does this support stream mode?

[qZhang88]

Does this support stream response? So I can continue accept partial message, and have nicer responding to user.

https://platform.openai.com/docs/api-reference/chat/create#chat/create-stream

[coin-au-carre]

Hello @qZhang88, yes it should be every field you see on the API reference documentation should be available.

[WenchiLe]

could anybody give an example?

[coin-au-carre]

I think you are right we have to implement this in order to make it work: https://stackoverflow.com/questions/31238626/curl-structuring-request-to-validate-server-sent-events

https://developer.mozilla.org/en-US/docs/Web/API/Server-sent_events/Using_server-sent_events#event_stream_format.

I will look into this unless someone is willing to take it

[AssetOverflow]

hi! just curious about the status of this one?

[WenchiLe]

I've written a demo using libcurl which may help.

class A
{
    public:
      virtual void print(std::string& s){ std::cout << "Class A: i:" << i << "s: " << s;i++;}
    private:
      int i = 1;
};

class B: public A
{
    public:
      virtual void print(std::string& s) override { std::cout << "Class B print: " << s <<endl; }
    private:
      int i = 2;
};

size_t WriteCallback(void* contents, size_t size, size_t nmemb, A* output) {
    size_t totalSize = size * nmemb;
    std::string s((char*)contents, totalSize);
    output->print(s);
    return totalSize;
}


void GetToOpenAIUsingCURL()
{
    CURL* curl;
    CURLcode res;

    B* a = new B();

    // 初始化libcurl
    curl_global_init(CURL_GLOBAL_DEFAULT);

    // 创建一个CURL对象
    curl = curl_easy_init();
    if (curl) {
        // 设置请求URL
        curl_easy_setopt(curl, CURLOPT_URL, "http://.../ai/proxy/v1/chat/completions");

        // 设置POST请求
        curl_easy_setopt(curl, CURLOPT_POST, 1L);

        // 设置请求体数据
        std::string postData = R"(
        {
            "model": "gpt-3.5-turbo",
            "messages":[{"role":"user", "content":"如何制作三明治"}],
            "max_tokens": 1000,
            "temperature": 0,
            "stream": true
        }
        )";
        curl_easy_setopt(curl, CURLOPT_POSTFIELDS, postData.c_str());

        // 设置请求头
        struct curl_slist* headers = NULL;
        headers = curl_slist_append(headers, "Content-Type: application/json");
        headers = curl_slist_append(headers, "Authorization: Bearer sk-...");
        headers = curl_slist_append(headers,  "Accept: text/event-stream");
        curl_easy_setopt(curl, CURLOPT_HTTPHEADER, headers);

        std::string response;
        curl_easy_setopt(curl, CURLOPT_WRITEFUNCTION, WriteCallback);
        curl_easy_setopt(curl, CURLOPT_WRITEDATA, a);

        // 发送请求
        std::cout << "curl_easy_perform start" << std::endl;
        res = curl_easy_perform(curl);
        std::cout << "curl_easy_perform end" << std::endl;
        this_thread::sleep_for(chrono::milliseconds(1000));
        // 检查请求是否成功
        if (res != CURLE_OK) {
            std::cerr << "curl_easy_perform() failed: " << curl_easy_strerror(res) << std::endl;
        } else {
            // 打印响应数据
            std::cout << "Response: " << response << std::endl;
        }

        // 释放请求头内存
        curl_slist_free_all(headers);

        // 清理CURL对象
        curl_easy_cleanup(curl);
    }

    // 清理libcurl
    curl_global_cleanup();
}

[coin-au-carre]

Sorry guys I will defintely take a look asap about this feature thanks for the example :)

[qZhang88]

Thanks hint from @WenchiLe

create two data type here, so I can poll stream chunk data outside from libcurl request.
Note: the stream data is not parsed yet, get_chunk and get_final will not work

struct StreamChunk {
    std::string text;
    bool        is_end;
};

class StreamResponse {
public:
    StreamResponse() : is_end_(false) {}

    void set(const std::string & chunk) {
        std::unique_lock<std::mutex> lk(mutex_);
        chunk_ += chunk;
        text_ += chunk;
        cv_.notify_all();
    }

    StreamChunk get_chunk() {
        std::unique_lock<std::mutex> lk(mutex_);
        cv_.wait(lk, [this] () { return chunk_ != ""; });
        std::string ret = chunk_;
        chunk_ = "";
        return {ret, is_end_};
    }

    std::string get_final() {
        std::unique_lock<std::mutex> lk(mutex_);
        cv_.wait(lk, [this] () { return is_end_ == true; });
        return text_;
    }

private:
    std::string chunk_;
    std::string text_;
    bool is_end_;

    std::mutex mutex_;
    std::condition_variable cv_;
};

struct CurlData{
    char* data;
    int64_t size;
};

right after static size_t writeFunction(void* ptr, size_t size, size_t nmemb, std::string* data), add s static variable and new WRITEFUNCTION.

    static StreamResponse * stream_response_;

    static size_t writeStreamFunction(void* ptr, size_t size, size_t nmemb, void * data) {
        size_t realsize = size * nmemb;
        std::string text((char*)ptr, realsize);
        stream_response_->set(text);
        return size * nmemb;
    }

pass in StreamResponse object response in ``

inline Response Session::makeRequest(StreamResponse& response, const std::string& contentType) {
    // ...
    stream_response_ = &response;
    curl_easy_setopt(curl_, CURLOPT_WRITEFUNCTION, writeStreamFunction);
    curl_easy_setopt(curl_, CURLOPT_WRITEDATA, nullptr);
    // ...

[adamakio]

Hey, just wondering if this feature will be added?

[psych0v0yager]

Is streaming ready to go?

It looks like @qZhang88 and @WenchiLe figured it out.

I don't mind contributing, what still needs to be done to implement streaming support?