Azure OpenAI Completion API返回非预期HTML代码问题求助

问题说明 #

调用Azure OpenAI Completion API提取职位描述核心意图时，返回了Laravel模板HTML代码，而非预期的纯文本列表。

具体解决步骤 #

1. 重构提示词，强化约束 #

修正原提示词拼写错误，明确禁止返回HTML、模板代码等非目标内容，强化格式要求：

You are a professional Hiring Manager. Extract the top 5 core intents from the following job description.
- Keep each intent concise (around 15 words), preserve original meaning, and ensure clarity.
- Format your response as 5 bullet points starting with "•", no other formatting allowed.
- **DO NOT return HTML, code templates, or any content not directly related to the intents.**
- Do not justify your answers, only include information from the provided job description.

Below is the job description:
{text_desc}

2. 调整API调用参数 #

• 降低temperature值至0.3-0.5，减少输出随机性，让模型更严格遵循指令
• 添加stop参数，截断无关内容的输出：

  response = openai.Completion.create(
      prompt=full_prompt,
      temperature=0.4,
      max_tokens=600,
      engine=deployment_name,
      stop=["</div>", "@endforeach", "@endsection"]
  )
  

3. 响应后处理过滤 #

在获取响应后，添加代码过滤HTML标签和无关内容，提取有效条目：

import re

# 获取原始响应
chatgpt_response = response['choices'][0].text.strip(" \n")

# 移除所有HTML标签
cleaned_response = re.sub(r'<.*?>', '', chatgpt_response)
# 提取以•开头的条目
valid_items = [item.strip() for item in cleaned_response.split('•') if item.strip()]
# 取前5条（如果有多余）
final_response = '\n• '.join(valid_items[:5])

4. 检查部署模型 #

确认Azure上部署的aidoos-openai模型版本是否适配，建议使用gpt-3.5-turbo-instruct（Completion API对应模型），避免使用训练数据包含大量模板代码的旧模型。

内容的提问来源于stack exchange，提问作者REHAMAN SHAIK