refactor(chat): AI-native prompts, remove interview heuristics

- Drop interview_reply_length and utterance_substance; always run stage LLM
  and memory retrieval when enabled; trim Settings fields and .env.example.
- Replace guided/opening prompts with compact fact blocks plus unified
  behavior guidance; slim background_voice and persona to tone hints.
- InterviewAgent uses fixed chat_interview max_tokens/chars/segments.

Also includes stacked work: profile followup/extract path, evaluation rubric
and judge schema updates, transcript SPLIT handling in execution service,
user export markdown split tests, and golden case fixture.
This commit is contained in:
Kevin
2026-04-06 22:22:50 +08:00
parent ca8bcc8489
commit 2fded6fbd9
27 changed files with 426 additions and 1349 deletions

View File

@@ -1,20 +1,24 @@
"""对话评审 rubric 文本v1"""
TURN_JUDGE_INSTRUCTIONS = """你是「岁月留书」访谈对话质量评审。根据下面维度给本轮 AI 回复打分0-100 为 total_score各子分上限已注明总和应合理)。
TURN_JUDGE_INSTRUCTIONS = """你是「岁月留书」访谈对话质量评审。根据下面维度给本轮 AI 回复打分0-100 为 total_score各子分上限已注明子分之和应与 total_score 大体一致)。
维度(参考):
- 情绪承接与共情emotion_score最高 30
- 信息获取与追问information_score最高 25
- 结构化访谈推进structure_score最高 15
- 提问质量question_score最高 15
- 人物理解与一致性persona_score最高 15
- 信息获取与追问information_score最高 20
- 结构化访谈推进structure_score最高 10
- 提问质量question_score最高 10
- 人物理解与一致性persona_score最高 10
- 重复抑制repetition_score最高 10是否重复了上 12 轮已问过的问题或同一资料槽;高度重复则低分
- 自然流畅naturalness_score最高 10是否像朋友聊天有无不必要采访腔、总结腔、流程感
输出 JSON**json** 字段名如下:
total_score, emotion_score, information_score, structure_score, question_score, persona_score, rationale
total_score, emotion_score, information_score, structure_score, question_score, persona_score, repetition_score, naturalness_score, rationale
只输出 JSON。"""
CONV_JUDGE_INSTRUCTIONS = """你是访谈整段对话评审。给定完整 transcript用户与 AI 多轮),打一个综合 total_score0-100,并给出 dimension_scores 对象(可为空对象),以及 rationale
CONV_JUDGE_INSTRUCTIONS = """你是访谈整段对话评审。给定完整 transcript用户与 AI 多轮),打一个综合 total_score0-100
dimension_scores 建议至少包含emotion, information, structure, repetition, naturalness各 0-100 相对分量即可),用于反映整段是否重复盘问、是否自然;另可有 rationale。
只输出 JSONtotal_score, dimension_scores, rationale。"""