Azure OpenAI Chat Playground与.NET API输出不一致问题求助

问题描述 #

在提取文本中多次提及主题的项目中，使用相同提示词、参数配置和模型部署，分别在Azure OpenAI Chat Playground和通过.NET应用调用GetChatCompletions()函数时，出现输出差异：

• Chat Playground稳定识别1个符合预期的主主题
• .NET代码识别出2-3个主题，其中部分仅在文本中提及一次，不符合项目目标

统一配置信息 #

• 参数：Temperature:0.2，MaxTokens:800，NucleusSamplingFactor:0.95，FrequencyPenalty:0，PresencePenalty:0，ChoiceCount:1
• 会话历史：两者均仅包含1条用户历史消息

提示词内容 #

You are a product review assistant that detects issues from customers' reviews and you will help product team summarize the prominent issues associated with the prodcuts by following these guidelines:
1. Review the provided text in XML format, specifically focusing on the 'review' tag.
2. Base your analysis strictly on the provided reviews, without including external information.
3. Keep your output concise and focused on the key details related to product problems.
4. If multiple prominent issues are found, provide separate summaries for each.
5. Consider a problem to be widespread only if it is mentioned in multiple reviews with similar keywords, patterns, and nature of the issue.
6. If more than 5 reviews are provided by users, only include issues mentioned in 5 or more reviews in the summary, excluding any isolated problems. If less than 5 reviews are provided, include all systemic issues regardless of their frequency in the summary.
7. Combine reviews about the same system component into a single entry to avoid redundancy.
8. Begin each issue with a bolded title summarizing the problem, followed by detailed information in this structured format with bolded category names:
- Symptom: Describe the issue users are experiencing.
- Workaround: Provide any user-found workarounds or solutions.
- Frequency Count: Count the number of times this issue is mentioned in the reviews.
If any above category is not mentioned in the given reviews, do not include the category name in the summary.

.NET核心代码 #

string endpoint = "xxxxxx";
string key = "yyyy";
OpenAIClient client = new OpenAIClient(new Uri(endpoint), new AzureKeyCredential(key));

public ChatCompletionsOptions GetSummary(string parameters, string deploymentName, string prompt, string input){

ChatCompletionsOptions chatCompletionsOptions = JsonConvert.DeserializeObject<ChatCompletionsOptions>(parameters);
chatCompletionsOptions.DeploymentName = deploymentName;
chatCompletionsOptions.Messages.Add(new ChatRequestSystemMessage(prompt));
chatCompletionsOptions.Messages.Add(new ChatRequestUserMessage(input));

// Call the OpenAI API to summarize the XML text
Response<ChatCompletions> response = _client.GetChatCompletions(chatCompletionsOptions);
}

可能原因 #

• Chat Playground隐式格式化：Playground会对提示词、用户消息做自动格式化（如去除多余空格、调整换行、添加隐式系统前缀），这些隐藏处理会影响模型理解逻辑。
• 参数序列化偏差：通过JsonConvert.DeserializeObject加载参数时，可能存在参数映射错误（如NucleusSamplingFactor与API的top_p是否完全对应）或缺失，导致实际调用参数与配置不符。
• 输入文本格式差异：Playground可能自动转义XML特殊字符或调整文本格式，而.NET代码直接传入原始XML，导致模型对输入的解析出现偏差。
• API版本不一致：Playground可能使用较新的API版本，而.NET SDK版本较旧，两者的API行为存在细微差异。

解决方案建议 #

1. 对比请求详情
   使用Fiddler或Azure Monitor捕获Playground和.NET代码的API请求，对比messages结构、参数配置、api-version请求头是否完全一致，重点检查系统提示、用户消息的格式和内容是否无差异。

2. 显式构建参数
   避免依赖反序列化加载参数，手动构建ChatCompletionsOptions并显式设置所有参数，确保与Playground配置完全匹配：

   var chatCompletionsOptions = new ChatCompletionsOptions
   {
       DeploymentName = deploymentName,
       Temperature = 0.2f,
       MaxTokens = 800,
       NucleusSamplingFactor = 0.95f,
       FrequencyPenalty = 0f,
       PresencePenalty = 0f,
       ChoiceCount = 1
   };
   

3. 统一输入文本处理
   对.NET代码中的XML输入做与Playground相同的格式化处理，比如去除多余空白、统一换行符、转义特殊字符，确保输入内容完全一致。

4. 同步API版本
   查看Playground请求详情中的api-version参数，将.NET SDK版本升级至对应版本，确保API行为一致。

5. 强化提示词约束
   在提示词中添加更明确的约束，比如在规则前补充：*注意：仅汇总在多个评论中重复出现的问题，孤立提及一次的问题必须排除*，减少模型歧义。

6. 设置固定种子值
   如果模型支持，设置Seed参数为固定值，确保生成结果可复现，便于调试对比。

内容的提问来源于stack exchange，提问作者Veronica