forestwanglin / openai-java Goto Github PK

这个同 https://github.com/TheoKanning/openai-java 功能和结构都没有多大差别啊
也都没有vision api

https://github.com/openai/openai-cookbook/blob/main/examples/How_to_count_tokens_with_tiktoken.ipynb这里只提到了普通的对话请求如何计算tokens，其余的tools以及function等请求计算tokens的依据是什么呢？

https://platform.openai.com/docs/api-reference/assistants
https://platform.openai.com/docs/guides/function-calling/parallel-function-calling

Hello,
According to the documentation, the Audio API is compatible. However, upon utilizing the Java API for Audio Transcription, I encounter an error stating "Unrecognized file format." I attempted this with various file formats such as m4a, mp3, mp4, and wav.

CreateAudioTranscriptionRequest request = CreateAudioTranscriptionRequest.builder().model("whisper-1") .filePath(filePath).language("en").responseFormat(ResponseFormat.TEXT).build(); AudioResponse response = new OpenAiService(token, Duration.ZERO).createAudioTranscription(request);

Exception in thread "main" xyz.felh.openai.OpenAiHttpException: Unrecognized file format. Supported formats: ['flac', 'm4a', 'mp3', 'mp4', 'mpeg', 'mpga', 'oga', 'ogg', 'wav', 'webm']
at xyz.felh.openai.OpenAiService.execute(OpenAiService.java:128)
at xyz.felh.openai.OpenAiService.createAudioTranscription(OpenAiService.java:461)
at openai.audio.SpeechToText.main(SpeechToText.java:40)
Caused by: retrofit2.adapter.rxjava3.HttpException: HTTP 400
at retrofit2.adapter.rxjava3.BodyObservable$BodyObserver.onNext(BodyObservable.java:57)
at retrofit2.adapter.rxjava3.BodyObservable$BodyObserver.onNext(BodyObservable.java:38)
at retrofit2.adapter.rxjava3.CallEnqueueObservable$CallCallback.onResponse(CallEnqueueObservable.java:62)
at retrofit2.OkHttpCall$1.onResponse(OkHttpCall.java:161)
at okhttp3.internal.connection.RealCall$AsyncCall.run(RealCall.kt:519)
at java.base/java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1136)
at java.base/java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:635)
at java.base/java.lang.Thread.run(Thread.java:840)

Is there any resolution available for this problem? Even a temporary workaround would be appreciated.

The token count for Tools is still having a difference of 1.
Request Body:

{
	"model": "gpt-4",
	"messages": [
		{
			"role": "assistant",
			"content": null,
			"tool_calls": [
				{
					"id": "call_Id8ycVMsW8gdsf7kSXfgAcf1",
					"type": "function",
					"function": {
						"name": "get_current_weather",
						"arguments": "{\n  \"location\": \"Boston, MA\"\n}"
					}
				}
			]
		},
		{
			"role": "tool",
			"tool_call_id": "call_Id8ycVMsW8gdsf7kSXfgAcf1",
			"name": "get_current_weather",
			"content": "29 degree celcius"
		}
	]
}

My count: 34
OpenAI Count: 35

My guess is that even "tool" role requires 3 tokens.

Thanks

使用Parallel function calling，会返回多个role为tool的message。
调用模型为gpt-4-1106-preview

第一次tool calling 先不讨论，因为一般人不会用流式调用第一次tool calls，

第二次tool calls调用prompt比官方多了1。
例子：

{
    "messages": [
        {
            "role": "system",
            "content": "你是一个数学达人"
        },
        {
            "role": "user",
            "content": "两道计算题 20+70; 30x100"
        },
        {
            "role": "assistant",
            "content": null,
            "tool_calls": [
                {
                    "id": "call_ouziXPZBrGmtdxh5BnBooKI3",
                    "type": "function",
                    "function": {
                        "name": "plus",
                        "arguments": "{\"numbers\": [20, 70]}"
                    }
                },
                {
                    "id": "call_jxP70EbkqL5L78EMVK76jFDh",
                    "type": "function",
                    "function": {
                        "name": "product",
                        "arguments": "{\"numbers\": [30, 100]}"
                    }
                }
            ]
        },
        {
            "tool_call_id": "call_ouziXPZBrGmtdxh5BnBooKI3",
            "role": "tool",
            "name": "plus",
            "content": "{\"numbers\": [20, 70], \"result\": \"90\"}"
        },
        {
            "tool_call_id": "call_jxP70EbkqL5L78EMVK76jFDh",
            "role": "tool",
            "name": "product",
            "content": "{\"numbers\": [30, 100], \"result\": \"30000\"}"
        }
    ]
}

OpenAI返回

"usage": {
    "prompt_tokens": 121,
    "completion_tokens": 29,
    "total_tokens": 150
  }

本库返回
prompt_tokens 122

与openai返回的token对比发现，几乎各chat模型都有计算方式错误或结果偏差，于是我自己从零建模和编写了token计算工具

你的部分token计算代码有严重错误，这里列举部分：

• vision base64读取
• vision图片缩放
• function call、functions
• tool call、tools
• 0301/0314与0613版本计算差异
• 多个参数（model、functions）组合使用对token计算的影响
• 应使用encodeOrdinary来跳过special tokens

由于精力有限我无法在开源代码上提交修改，本issue只是告知绝大部分token计算都有误，请你自己有精力时研究下

您好大佬,在maven仓库找不到最新的版本,请问在哪可以找到?

Great project i can't use because i use Java 8:

java: cannot access xyz.felh.openai.chat.ChatMessage
bad class file: lib/jtokkit/core-4.0.2024080901.jar!/xyz/felh/openai/chat/ChatMessage.class
class file has wrong version 65.0, should be 52.0
Please remove or make sure it appears in the correct subdirectory of the classpath.

Thanks for implementing token count logic in Java, this is very helpful!

I found a few cases in which the count calculated was different from the actual OpenAI count. While I know this is an estimation, it would be good if you could check the logic.

1. When an integer enum is passed as a function parameter, 1 less token is counted.
2. When the message contains role assistant with tool_calls and a subsequent message with role tool, 6 extra tokens are counted.

If you want, I can provide examples for the above.
Thanks!

感谢

when I count the number of tokens, this message is cut off after formatting and is not fully counted

{
            "role": "tool",
            "content": "{\"items\":[{\"title\":\"#MINSK NIGHTLIFE TOUR / BELARUS AFTER SANCTIONS JUNE ...\", \"link\":\"https://www.youtube.com/watch?v=d6te-xePaj0\", \"snippet\":\"Jun 26, 2022 ... Anfisa BELARUS•66K views · 16:51 · Go to channel · MANHATTAN NIGHTLIFE AREAS - PACKED BARS \u0026 CLUBS Summer Update【ENTIRE TOUR】Best ...\"}, {\"title\":\"THE 10 BEST Nightlife Activities in Minsk (Updated 2024) - Tripadvisor\", \"link\":\"https://www.tripadvisor.com/Attractions-g294448-Activities-c20-Minsk.html\", \"snippet\":\"Results 1 - 30 of 77 ... These places are best for nightlife in Minsk: RetravelMe Belarus · Party Bus · HookahPlace Yakuba Kolasa · Nuahule Krasnaya · Dictator Bar.\"}, {\"title\":\"MINSK AT NIGHT WITH @IrishPartizan - YouTube\", \"link\":\"https://www.youtube.com/watch?v=1siveAZHym0\", \"snippet\":\"Aug 5, 2023 ... scene, this video is your ultimate guide to the hottest night spots in Minsk ... MINSK NIGHTLIFE AFTER SANCTIONS / BEST 5 NIGHT BARS IN MINSK.\"}]}",
            "name": "http",
            "tool_call_id": "call_ZPSUPvPgBhNMSzI2mXDW36XZ"
        }

public static String formatArguments(String arguments) {
        List<String> lines = new ArrayList<>();
        lines.add("{");
        JSONObject jsonObject = JSONObject.parseObject(arguments);
        List<String> properties = new ArrayList<>();
        for (String fieldName : jsonObject.keySet()) {
            properties.add(String.format("\"%s\":%s", fieldName, formatValue(jsonObject.get(fieldName))));
        }
        lines.add(String.join(",\n", properties));
        lines.add("}");
        return String.join("\n", lines);
    }

Output:
"{
"items":["","",""]
}"

大佬你好，今天看文档openai的submit_tool_outputs接口有stream输出设置，计划什么时候纳入开发

openai api：
https://api.openai.com/v1/threads/{thread_id}/runs/{run_id}/submit_tool_outputs