plan_classifier.py 4.0 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105
  1. """plan_classifier.py
  2. 用于鉴定上传的 Markdown 文件是否为训练计划。
  3. 采用双层过滤架构:
  4. 1. Level 1 - 关键字快筛(正则表达式 / 词频)
  5. 2. Level 2 - LLM (Lite) 语义鉴定
  6. """
  7. import os
  8. import json
  9. import re
  10. from typing import Any
  11. LLM_LITE_API_KEY: str | None = os.getenv("LLM_LITE_API_KEY") or os.getenv("LLM_API_KEY")
  12. LLM_LITE_BASE_URL: str = (os.getenv("LLM_LITE_BASE_URL") or "https://api.deepseek.com").rstrip("/")
  13. LLM_LITE_MODE: str = os.getenv("LLM_LITE_MODE_ID", "deepseek-chat")
  14. # 关键词仅用于快速排除明显无关的文件;不得据此直接接收文件。
  15. TRAINING_KEYWORDS = [
  16. "组间歇", "重量", "卧推", "硬拉", "深蹲", "划船", "哑铃", "杠铃", "RM", "组数",
  17. "力竭", "热身", "拉伸", "核心肌群", "训练", "计划", "运动", "健身", "锻炼", "动作",
  18. "次", "分钟",
  19. ]
  20. def _level1_keyword_check(text: str) -> tuple[bool, str] | None:
  21. """仅拒绝不含任何训练内容的文件,其余全部交给语义鉴定。"""
  22. if not any(keyword.lower() in text.lower() for keyword in TRAINING_KEYWORDS):
  23. return False, "完全不包含任何运动、健身相关的常见词汇"
  24. return None
  25. def _extract_json(text: str) -> dict[str, Any] | None:
  26. md_match = re.search(r"```(?:json)?\s*(\{[\s\S]+?\})\s*```", text)
  27. if md_match:
  28. try:
  29. return json.loads(md_match.group(1))
  30. except json.JSONDecodeError:
  31. pass
  32. brace_match = re.search(r"\{[\s\S]+?\}", text)
  33. if brace_match:
  34. try:
  35. return json.loads(brace_match.group(0))
  36. except json.JSONDecodeError:
  37. pass
  38. return None
  39. def _level2_llm_check(text: str) -> tuple[bool, str]:
  40. if not LLM_LITE_API_KEY:
  41. return False, "未配置 LLM API KEY 且无法通过规则验证"
  42. system_prompt = (
  43. "判断以下 Markdown 文本是否为一份可执行的健身/运动训练计划。\n"
  44. "仅仅提及动作名、重量、训练感受、健身知识、训练日志、复盘或文章,不是训练计划。\n"
  45. "只有包含面向未来执行的明确训练安排(例如动作、次数/时长/强度、顺序或频率)时才是训练计划。\n"
  46. "返回严格 JSON 格式:\n"
  47. "{\n"
  48. ' "is_plan": true 或 false,\n'
  49. ' "reason": "一句话说明理由"\n'
  50. "}"
  51. )
  52. try:
  53. import requests
  54. url = f"{LLM_LITE_BASE_URL}/chat/completions"
  55. headers = {
  56. "Authorization": f"Bearer {LLM_LITE_API_KEY}",
  57. "Content-Type": "application/json",
  58. }
  59. body = {
  60. "model": LLM_LITE_MODE,
  61. "messages": [
  62. {"role": "system", "content": system_prompt},
  63. {"role": "user", "content": text[:2000]} # 截断防止超长
  64. ],
  65. "max_tokens": 200,
  66. "temperature": 0.0,
  67. "response_format": {"type": "json_object"}
  68. }
  69. resp = requests.post(url, headers=headers, json=body, timeout=10)
  70. resp.raise_for_status()
  71. raw_text = resp.json()["choices"][0]["message"]["content"]
  72. parsed = _extract_json(raw_text)
  73. if parsed is None:
  74. return False, "LLM 返回无法解析"
  75. return bool(parsed.get("is_plan", False)), str(parsed.get("reason", ""))
  76. except Exception as exc:
  77. return False, f"LLM 调用异常: {exc}"
  78. def validate_training_plan(
  79. text: str, *, allow_external_models: bool = True
  80. ) -> dict[str, Any]:
  81. """主入口
  82. 返回: {"is_plan": bool, "reason": str, "stage": str}
  83. """
  84. res1 = _level1_keyword_check(text)
  85. if res1 is not None:
  86. return {"is_plan": res1[0], "reason": res1[1], "stage": "level1_keywords"}
  87. if not allow_external_models:
  88. return {
  89. "is_plan": False,
  90. "reason": "已关闭外部模型,无法对模糊的计划内容进行语义鉴定",
  91. "stage": "external_models_disabled",
  92. }
  93. res2 = _level2_llm_check(text)
  94. return {"is_plan": res2[0], "reason": res2[1], "stage": "level2_llm"}