reflexion.py 7.8 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223
  1. """Reflexion 自校验机制:关键步骤后自检质量,低于阈值触发针对性补抽/重试
  2. 核心职责:
  3. - self_check:让模型对照上下文检查结果质量,输出置信度与问题
  4. - reflect_and_retry:自校验 + 针对性重试(非整篇重跑),最多重试 2 次
  5. - 记录所有自校验置信度,提供统计接口
  6. """
  7. from __future__ import annotations
  8. import json
  9. import logging
  10. from typing import Callable
  11. from backend.config import settings
  12. from backend.agents.hello_agents_runner import build_reviewer_agent, run_reviewer_check
  13. from backend.llm.client import DeepSeekClient
  14. from backend.llm.json_utils import parse_json_safe
  15. logger = logging.getLogger(__name__)
  16. # 审核 Agent 的系统提示词
  17. REVIEWER_SYSTEM_PROMPT = "你是质量审核员,负责对照上下文检查 Agent 输出质量。只输出 JSON。"
  18. # 最多重试次数
  19. MAX_RETRIES = 2
  20. class Reflexion:
  21. """Reflexion 自校验机制
  22. 在关键步骤(抽取/Gap/创新)后启用:
  23. 1. 对照上下文自检结果质量,输出置信度;
  24. 2. 置信度低于阈值时触发针对性补抽/重试(retry_func),非整篇重跑;
  25. 3. 最多重试 MAX_RETRIES 次,仍不达标则标记 needs_human_review。
  26. """
  27. def __init__(self, client: DeepSeekClient) -> None:
  28. self.client = client
  29. # 记录所有自校验的置信度
  30. self.confidences: list[float] = []
  31. async def self_check(
  32. self,
  33. step_name: str,
  34. result: dict,
  35. context: str = "",
  36. ) -> dict:
  37. """让模型对照上下文检查结果质量
  38. Args:
  39. step_name: 步骤名称(如 Extractor / GapAnalyzer / InnovationGenerator)
  40. result: 步骤执行结果 dict
  41. context: 校验上下文(依赖步骤的结果或原文摘要)
  42. Returns:
  43. {
  44. "confidence": float, # 0-1 置信度
  45. "issues": list[str], # 发现的问题
  46. "needs_retry": bool, # 是否需要重试
  47. "retry_fields": list[str], # 需重试的字段或方面
  48. }
  49. """
  50. # 序列化结果,截断避免超长
  51. if isinstance(result, dict):
  52. result_str = json.dumps(result, ensure_ascii=False, default=str)[:6000]
  53. else:
  54. result_str = str(result)[:6000]
  55. context_str = context[:4000] if context else "(无额外上下文)"
  56. prompt = (
  57. f"请对照上下文检查以下步骤的结果质量。\n\n"
  58. f"步骤:{step_name}\n"
  59. f"上下文:\n{context_str}\n\n"
  60. f"结果:\n{result_str}\n\n"
  61. f"检查要求:\n"
  62. f"1. 评估结果是否准确、完整、无臆造;\n"
  63. f"2. 输出 0-1 的置信度(confidence);\n"
  64. f"3. 列出发现的问题(issues);\n"
  65. f"4. 置信度低于 {settings.CONFIDENCE_THRESHOLD} 时 needs_retry=true;\n"
  66. f"5. 指出需重试的字段或方面(retry_fields)。\n\n"
  67. f"严格以 JSON 输出,格式:\n"
  68. f'{{"confidence": 0.85, "issues": ["问题1"], '
  69. f'"needs_retry": false, "retry_fields": []}}'
  70. )
  71. # 通过 HelloAgents 框架 SimpleAgent(质量审核员角色)执行自校验
  72. try:
  73. reviewer = build_reviewer_agent(
  74. self.client.llm, system_prompt=REVIEWER_SYSTEM_PROMPT
  75. )
  76. content = await run_reviewer_check(reviewer, prompt)
  77. data = parse_json_safe(content)
  78. if not isinstance(data, dict):
  79. data = {}
  80. except Exception as exc:
  81. logger.warning("自校验 LLM 调用失败: %s", exc)
  82. data = {}
  83. # 解析置信度
  84. try:
  85. confidence = float(data.get("confidence", 0.0))
  86. except (TypeError, ValueError):
  87. confidence = 0.0
  88. confidence = max(0.0, min(1.0, confidence))
  89. # 解析是否需要重试
  90. needs_retry = data.get("needs_retry", False)
  91. if not isinstance(needs_retry, bool):
  92. needs_retry = bool(needs_retry)
  93. # 置信度低于阈值强制标记重试
  94. if confidence < settings.CONFIDENCE_THRESHOLD:
  95. needs_retry = True
  96. # 解析问题列表
  97. issues = data.get("issues", [])
  98. if not isinstance(issues, list):
  99. issues = [str(issues)] if issues else []
  100. # 解析重试字段
  101. retry_fields = data.get("retry_fields", [])
  102. if not isinstance(retry_fields, list):
  103. retry_fields = [str(retry_fields)] if retry_fields else []
  104. # 记录置信度
  105. self.confidences.append(confidence)
  106. check_result = {
  107. "confidence": confidence,
  108. "issues": issues,
  109. "needs_retry": needs_retry,
  110. "retry_fields": retry_fields,
  111. }
  112. logger.info(
  113. "步骤 %s 自校验: confidence=%.2f, needs_retry=%s",
  114. step_name,
  115. confidence,
  116. needs_retry,
  117. )
  118. return check_result
  119. async def reflect_and_retry(
  120. self,
  121. step_name: str,
  122. result: dict,
  123. context: str,
  124. retry_func: Callable,
  125. ) -> dict:
  126. """自校验 + 针对性重试
  127. 流程:
  128. 1. 调用 self_check 检查结果质量;
  129. 2. 若 needs_retry,调用 retry_func 进行针对性补抽/重试(非整篇重跑);
  130. 3. 最多重试 MAX_RETRIES 次;
  131. 4. 仍不达标则标记 needs_human_review 并返回当前结果。
  132. Args:
  133. step_name: 步骤名称
  134. result: 原始结果
  135. context: 校验上下文
  136. retry_func: 重试函数(async,无参,返回新的 result dict)
  137. Returns:
  138. 更新后的 result dict,附加 reflexion 字段与可能的 needs_human_review
  139. """
  140. current = result
  141. check = await self.self_check(step_name, current, context)
  142. retries = 0
  143. while check["needs_retry"] and retries < MAX_RETRIES:
  144. retries += 1
  145. logger.info(
  146. "步骤 %s 自校验未达标 (confidence=%.2f),触发针对性重试 %d/%d",
  147. step_name,
  148. check["confidence"],
  149. retries,
  150. MAX_RETRIES,
  151. )
  152. try:
  153. current = await retry_func()
  154. # 重试后重新自校验
  155. check = await self.self_check(step_name, current, context)
  156. except Exception as exc:
  157. logger.warning("步骤 %s 重试失败: %s", step_name, exc)
  158. break
  159. # 仍不达标,标记需人工复核
  160. needs_human_review = check["needs_retry"]
  161. if needs_human_review:
  162. logger.warning(
  163. "步骤 %s 经 %d 次重试仍未达标 (confidence=%.2f),标记需人工复核",
  164. step_name,
  165. retries,
  166. check["confidence"],
  167. )
  168. # 将自校验结果附加到 result
  169. if isinstance(current, dict):
  170. current["reflexion"] = check
  171. if needs_human_review:
  172. current["needs_human_review"] = True
  173. return current
  174. return {
  175. "result": current,
  176. "reflexion": check,
  177. "needs_human_review": needs_human_review,
  178. }
  179. def get_confidence_stats(self) -> dict:
  180. """返回置信度统计
  181. Returns:
  182. {avg_confidence, min_confidence, count}
  183. """
  184. if not self.confidences:
  185. return {"avg_confidence": 0.0, "min_confidence": 0.0, "count": 0}
  186. return {
  187. "avg_confidence": round(sum(self.confidences) / len(self.confidences), 4),
  188. "min_confidence": round(min(self.confidences), 4),
  189. "count": len(self.confidences),
  190. }