add groq support

Author: LeafCreeper

.github/workflows/check.yml

@@ -94,6 +94,7 @@ jobs:
           UISPsw: ${{ secrets.UISPSW }}
           COURSE_IDS: ${{ secrets.COURSE_IDS }}
           DASHSCOPE_API_KEY: ${{ secrets.DASHSCOPE_API_KEY }}
+          GROQ_API_KEY: ${{ secrets.GROQ_API_KEY }}
           SMTP_EMAIL: ${{ secrets.SMTP_EMAIL }}
           SMTP_PASSWORD: ${{ secrets.SMTP_PASSWORD }}
           RECEIVER_EMAIL: ${{ secrets.RECEIVER_EMAIL }}

.github/workflows/single_run.yml

@@ -95,6 +95,7 @@ jobs:
           UISPsw: ${{ secrets.UISPSW }}
           COURSE_IDS: ${{ inputs.course_ids }}
           DASHSCOPE_API_KEY: ${{ secrets.DASHSCOPE_API_KEY }}
+          GROQ_API_KEY: ${{ secrets.GROQ_API_KEY }}
           SMTP_EMAIL: ${{ secrets.SMTP_EMAIL }}
           SMTP_PASSWORD: ${{ secrets.SMTP_PASSWORD }}
           RECEIVER_EMAIL: ${{ secrets.RECEIVER_EMAIL }}

src/config.py

@@ -26,10 +26,18 @@
     "ZhipuAI/GLM-5",
     "deepseek-ai/DeepSeek-V3.2",
     "MiniMax/MiniMax-M2.5",
-    "ZhipuAI/GLM-4.7",
-    "deepseek-ai/DeepSeek-V3-0324",
     "Qwen/Qwen3.5-397B-A17B",
-    "Qwen/Qwen2.5-72B-Instruct"
+    "ZhipuAI/GLM-4.7"
+]
+
+# Groq fallback (for content policy bypass)
+GROQ_API_KEY = os.environ.get("GROQ_API_KEY", "")
+GROQ_BASE_URL = "https://api.groq.com/openai/v1"
+GROQ_MODELS = [
+    "llama-3.3-70b-versatile",
+    "llama-4-maverick",
+    "llama-4-scout",
+    "openai/gpt-oss-120b",
 ]
 
 # QQ SMTP

src/summarizer.py

@@ -8,18 +8,24 @@
 
 SYSTEM_PROMPT = (
 "你是一个专业的课程助教。你的任务是根据用户提供的课程录音文本，生成用于学生自学和期末复习的详细结构化笔记。\n"
-"1. **直接输出**：不要包含任何“好的”、“没问题”、“以下是总结”等客套话。不要输出全局课程名称大标题（由系统自动生成），直接开始总结即可。\n"
-"2. **文本清洗**：语言必须通顺、逻辑清晰，严格去除口语化表达（如“呃”、“啊”、“那么”）、重复句和无意义的录音识别错误等。\n"
+"1. **直接输出**：不要包含任何"好的"、"没问题"、"以下是总结"等客套话。不要输出全局课程名称大标题（由系统自动生成），直接开始总结即可。\n"
+"2. **文本清洗**：语言必须通顺、逻辑清晰，严格去除口语化表达（如"呃"、"啊"、"那么"）、重复句和无意义的录音识别错误等。\n"
 "3. **格式严格**：\n"
    "- 必须使用 Markdown 格式排版。\n"
   " - **标题级别限制**：只允许使用三级及更低级别的标题（即只能使用 `###`、`####`、`#####`），禁止使用 `#` 和 `##`。\n"
   "- 合理使用加粗、列表、表格来组织信息，确保结构清晰。\n"
 "4. **公式规范**：所有数学公式或科学变量必须使用规范的 LaTeX 语法（行内公式用 $...$，行间公式用 $$...$$）。\n"
-"5. **忠于原文与详尽**：总结必须尽可能详细且长，包含具体的推导细节、案例、文献或者核心概念，不要过度概括。禁止捏造录音中未提及的内容。\n"
+"5. **忠于原文与详尽**：总结必须尽可能详细且足够长，包含具体的推导细节、案例、文献或者核心概念，不要过度概括。禁止捏造录音中未提及的内容。\n"
 "6. 你需要格外注意课程中是否提及了作业、考试、签到、组队等关键事项，如果有的话，用三级标题【课程事项提醒】标注在开头。"
 )
 
 
+def _is_content_blocked(error: Exception) -> bool:
+    """Check if an API error is a content policy rejection."""
+    msg = str(error).lower()
+    return "data_inspection_failed" in msg or "inappropriate content" in msg
+
+
 class Summarizer:
     """Course lecture summarizer using ModelScope OpenAI-compatible API."""
 
@@ -32,12 +38,41 @@ def __init__(self):
         )
         self.models = list(config.LLM_MODELS)
 
+        self._groq_client = None
+        if config.GROQ_API_KEY:
+            self._groq_client = OpenAI(
+                api_key=config.GROQ_API_KEY,
+                base_url=config.GROQ_BASE_URL,
+            )
+
+    def _call_llm(self, client: OpenAI, model: str,
+                  title: str, content: str) -> str:
+        """Send a summarization request to a single model. Returns result text."""
+        t0 = time.time()
+        response = client.chat.completions.create(
+            model=model,
+            messages=[
+                {"role": "system", "content": SYSTEM_PROMPT},
+                {
+                    "role": "user",
+                    "content": f"以下是课程《{title}》的录音文本，请总结：\n\n{content}",
+                },
+            ],
+            temperature=0.3,
+            timeout=120,
+        )
+        result = response.choices[0].message.content
+        elapsed = time.time() - t0
+        print(
+            f"[Summarizer] Done ({model}): {len(content)} chars input"
+            f" → {len(result)} chars output in {elapsed:.0f}s"
+        )
+        return result
+
     def summarize(self, title: str, content: str) -> tuple[str, str]:
         """Summarize lecture content, trying multiple models on failure.
 
-        Args:
-            title: Lecture title for context.
-            content: Full transcript text.
+        If all primary models fail due to content policy, falls back to Groq.
 
         Returns:
             (summary, model_used) tuple.
@@ -46,31 +81,31 @@ def summarize(self, title: str, content: str) -> tuple[str, str]:
             return ("（内容为空）", "")
 
         errors = []
+        content_blocked = False
+
+        # Primary: ModelScope models
         for model in self.models:
             try:
-                t0 = time.time()
-                response = self.client.chat.completions.create(
-                    model=model,
-                    messages=[
-                        {"role": "system", "content": SYSTEM_PROMPT},
-                        {
-                            "role": "user",
-                            "content": f"以下是课程《{title}》的录音文本，请总结：\n\n{content}",
-                        },
-                    ],
-                    temperature=0.3,
-                    timeout=120,
-                )
-                result = response.choices[0].message.content
-                elapsed = time.time() - t0
-                print(
-                    f"[Summarizer] Done ({model}): {len(content)} chars input"
-                    f" → {len(result)} chars output in {elapsed:.0f}s"
-                )
+                result = self._call_llm(self.client, model, title, content)
                 return (result, model)
             except Exception as e:
                 print(f"[Summarizer] {model} failed: {type(e).__name__}: {e}")
                 errors.append(f"{model}: {e}")
+                if _is_content_blocked(e):
+                    content_blocked = True
+
+        # Fallback: Groq models (when content policy blocks primary models)
+        if content_blocked and self._groq_client:
+            print("[Summarizer] Content blocked by primary platform, trying Groq...")
+            for model in config.GROQ_MODELS:
+                try:
+                    result = self._call_llm(
+                        self._groq_client, model, title, content,
+                    )
+                    return (result, f"groq/{model}")
+                except Exception as e:
+                    print(f"[Summarizer] groq/{model} failed: {type(e).__name__}: {e}")
+                    errors.append(f"groq/{model}: {e}")
 
         raise RuntimeError(
             "All LLM models failed:\n" + "\n".join(errors)