6772e1269c
Add memoir_readiness_service and router tests; extend judge schemas/services, replay_service, and conversation rubric; align story route agent, payload, prompts, and story_pipeline_sync; update agent logging, config, and DI. Document internal-eval; add replayDraft util and PlaygroundPage changes in app-eval-web.
334 lines
14 KiB
Python
334 lines
14 KiB
Python
"""评审 LLM 结构化输出（json_object）。
|
||
|
||
成稿（回忆录）子项上限已自洽为 **总分 100**（由原 110 分表等比例收紧整数档，见附件 rubric）。
|
||
"""
|
||
|
||
from __future__ import annotations
|
||
|
||
from typing import Any, Self
|
||
|
||
from pydantic import BaseModel, ConfigDict, Field, model_validator
|
||
|
||
|
||
class JudgeEvidenceRef(BaseModel):
|
||
"""评审引用：便于人工复核（对话 Turn 或成稿片段定位）。"""
|
||
|
||
model_config = ConfigDict(extra="ignore")
|
||
|
||
dimension: str = ""
|
||
turn_index: int = Field(default=-1, ge=-1)
|
||
snippet: str = Field(default="", max_length=400)
|
||
|
||
|
||
def _is_judge_list_placeholder_empty(s: str) -> bool:
|
||
"""LLM 有时输出单句占位（如 'None identified.'）而非 JSON 数组，按空列表处理。"""
|
||
t = s.strip()
|
||
if not t:
|
||
return True
|
||
tl = t.lower().rstrip(".")
|
||
if tl in (
|
||
"none",
|
||
"none identified",
|
||
"n/a",
|
||
"na",
|
||
"-",
|
||
"nil",
|
||
"null",
|
||
"no issues",
|
||
"no issue",
|
||
"not applicable",
|
||
):
|
||
return True
|
||
tc = t.rstrip("。")
|
||
if tc in ("无", "暂无", "未发现", "没有"):
|
||
return True
|
||
return False
|
||
|
||
|
||
def _coerce_judge_str_list(value: Any) -> list[Any]:
|
||
"""将评审 JSON 中的 list[str] 字段从 str / null 规范为列表（兼容 GLM-5 等输出的非数组形态）。"""
|
||
if value is None:
|
||
return []
|
||
if isinstance(value, list):
|
||
return value
|
||
if isinstance(value, str):
|
||
s = value.strip()
|
||
if _is_judge_list_placeholder_empty(s):
|
||
return []
|
||
return [s]
|
||
return []
|
||
|
||
|
||
class TurnJudgeOutput(BaseModel):
|
||
"""单轮 / 整段对话质量（情绪强化版 100 分，15 个细项）。"""
|
||
|
||
model_config = ConfigDict(extra="ignore")
|
||
|
||
# 一、情绪价值与陪伴感（30）
|
||
emotion_carry: float = Field(ge=0, le=10, description="情绪承接能力")
|
||
empathy_depth: float = Field(ge=0, le=8, description="共情深度")
|
||
emotion_safety: float = Field(ge=0, le=6, description="情绪安全感")
|
||
emotion_guidance: float = Field(ge=0, le=6, description="情绪引导能力")
|
||
|
||
# 二、信息获取能力（25）
|
||
fact_mining: float = Field(ge=0, le=8, description="关键事实挖掘")
|
||
info_completeness_guide: float = Field(ge=0, le=8, description="信息完整性引导")
|
||
info_depth_mining: float = Field(ge=0, le=9, description="信息深度挖掘")
|
||
|
||
# 三、人物建模能力（15）
|
||
persona_understanding: float = Field(ge=0, le=7, description="人物理解")
|
||
persona_consistency_verify: float = Field(ge=0, le=4, description="人物一致性验证")
|
||
persona_expression_guide: float = Field(ge=0, le=4, description="人物表达引导")
|
||
|
||
# 四、结构化引导（15）
|
||
interview_structure: float = Field(ge=0, le=6, description="访谈结构")
|
||
context_memory: float = Field(ge=0, le=5, description="上下文记忆")
|
||
rhythm_control: float = Field(ge=0, le=4, description="节奏控制")
|
||
|
||
# 五、提问质量（15）
|
||
question_quality: float = Field(ge=0, le=7, description="问题质量")
|
||
follow_up_depth: float = Field(ge=0, le=5, description="追问能力")
|
||
non_leading: float = Field(ge=0, le=3, description="非引导性")
|
||
|
||
total_score: float = Field(ge=0, le=100)
|
||
rationale: str = ""
|
||
|
||
major_strengths: list[str] = Field(default_factory=list)
|
||
major_issues: list[str] = Field(default_factory=list)
|
||
insufficient_evidence: list[str] = Field(default_factory=list)
|
||
evidence_refs: list[JudgeEvidenceRef] = Field(default_factory=list)
|
||
confidence: float = Field(default=0.75, ge=0.0, le=1.0)
|
||
|
||
# 与历史 JSON 对齐的一级聚合分（由细项派生，可缺省由模型填写）
|
||
emotion_score: float = Field(default=0, ge=0, le=30)
|
||
information_score: float = Field(default=0, ge=0, le=25)
|
||
persona_score: float = Field(default=0, ge=0, le=15)
|
||
structure_score: float = Field(default=0, ge=0, le=15)
|
||
question_score: float = Field(default=0, ge=0, le=15)
|
||
|
||
@model_validator(mode="before")
|
||
@classmethod
|
||
def _coerce_null_lists(cls, data: Any) -> Any:
|
||
if isinstance(data, dict):
|
||
for key in ("major_strengths", "major_issues", "insufficient_evidence"):
|
||
data[key] = _coerce_judge_str_list(data.get(key))
|
||
if data.get("evidence_refs") is None:
|
||
data["evidence_refs"] = []
|
||
return data
|
||
|
||
@model_validator(mode="after")
|
||
def _cap_meta_fields_and_sync_totals(self) -> Self:
|
||
def _cap_str_list(xs: list[str], *, max_items: int, max_chars: int) -> list[str]:
|
||
out: list[str] = []
|
||
for x in xs[:max_items]:
|
||
s = str(x).strip()
|
||
if s:
|
||
out.append(s[:max_chars])
|
||
return out
|
||
|
||
object.__setattr__(
|
||
self,
|
||
"major_strengths",
|
||
_cap_str_list(self.major_strengths, max_items=8, max_chars=200),
|
||
)
|
||
object.__setattr__(
|
||
self,
|
||
"major_issues",
|
||
_cap_str_list(self.major_issues, max_items=10, max_chars=200),
|
||
)
|
||
object.__setattr__(
|
||
self,
|
||
"insufficient_evidence",
|
||
_cap_str_list(self.insufficient_evidence, max_items=10, max_chars=200),
|
||
)
|
||
refs = list(self.evidence_refs)[:12]
|
||
object.__setattr__(self, "evidence_refs", refs)
|
||
|
||
emotion = (
|
||
self.emotion_carry
|
||
+ self.empathy_depth
|
||
+ self.emotion_safety
|
||
+ self.emotion_guidance
|
||
)
|
||
information = (
|
||
self.fact_mining + self.info_completeness_guide + self.info_depth_mining
|
||
)
|
||
persona = (
|
||
self.persona_understanding
|
||
+ self.persona_consistency_verify
|
||
+ self.persona_expression_guide
|
||
)
|
||
structure = self.interview_structure + self.context_memory + self.rhythm_control
|
||
question = self.question_quality + self.follow_up_depth + self.non_leading
|
||
expected = emotion + information + persona + structure + question
|
||
# 细项为唯一事实来源：LLM 常把 total_score 写成 100 与前四项打满但情绪块少 1 分等情况不一致
|
||
synced = max(0.0, min(100.0, round(float(expected), 2)))
|
||
object.__setattr__(self, "total_score", synced)
|
||
object.__setattr__(self, "emotion_score", emotion)
|
||
object.__setattr__(self, "information_score", information)
|
||
object.__setattr__(self, "persona_score", persona)
|
||
object.__setattr__(self, "structure_score", structure)
|
||
object.__setattr__(self, "question_score", question)
|
||
return self
|
||
|
||
|
||
# 整条 transcript 与单轮使用同一套细项
|
||
ConversationJudgeOutput = TurnJudgeOutput
|
||
|
||
|
||
class MemoirJudgeOutput(BaseModel):
|
||
"""成稿回忆录评分（总分 100，子项上限见 rubric）。"""
|
||
|
||
model_config = ConfigDict(extra="ignore")
|
||
|
||
# 一、真实性与覆盖（小计最高 23；由原 25 收紧）
|
||
mem_fidelity: float = Field(ge=0, le=9, description="记忆忠实度")
|
||
mem_factual_accuracy: float = Field(ge=0, le=5, description="事实准确性")
|
||
mem_factual_coverage: float = Field(ge=0, le=5, description="事实覆盖率")
|
||
mem_traceability: float = Field(ge=0, le=4, description="记忆可追溯性")
|
||
|
||
# 二、信息质量（小计最高 14；由原 15 收紧）
|
||
info_slot_coverage: float = Field(ge=0, le=6, description="槽位覆盖度")
|
||
info_sufficiency: float = Field(ge=0, le=4, description="信息充分性")
|
||
info_density: float = Field(ge=0, le=4, description="信息密度")
|
||
|
||
# 三、叙事结构（小计最高 14；由原 15 收紧）
|
||
narr_structure: float = Field(ge=0, le=6, description="故事结构")
|
||
narr_paragraphs: float = Field(ge=0, le=5, description="段落组织")
|
||
narr_pacing: float = Field(ge=0, le=3, description="节奏控制")
|
||
|
||
# 四、语言与文笔（小计最高 18；由原 20 及六项上限一并收紧）
|
||
lang_fluency: float = Field(ge=0, le=3, description="语言流畅度")
|
||
lang_conciseness: float = Field(ge=0, le=3, description="表达精炼度")
|
||
lang_literary: float = Field(ge=0, le=4, description="文笔质量")
|
||
lang_controlled_expansion: float = Field(ge=0, le=4, description="控制性扩写能力")
|
||
lang_detail: float = Field(ge=0, le=2, description="细节还原与强化")
|
||
lang_style: float = Field(ge=0, le=2, description="风格一致性")
|
||
|
||
# 五、情感表达（小计最高 9；由原 10 收紧）
|
||
emo_authenticity: float = Field(ge=0, le=5, description="情感真实度")
|
||
emo_depth: float = Field(ge=0, le=4, description="情感深度")
|
||
|
||
# 六、人物建模（小计最高 9；由原 10 收紧）
|
||
char_understanding: float = Field(ge=0, le=4, description="人物理解")
|
||
char_consistency: float = Field(ge=0, le=3, description="人物一致性")
|
||
char_integration: float = Field(ge=0, le=2, description="人物融入度")
|
||
|
||
# 七、连贯性（小计最高 4；由原 5 收紧）
|
||
coh_timeline: float = Field(ge=0, le=2, description="时间线一致性")
|
||
coh_cross_chapter: float = Field(ge=0, le=2, description="跨章节关联")
|
||
|
||
# 八、表达丰富度（小计最高 5）
|
||
rich_analogy: float = Field(ge=0, le=3, description="类比与引用")
|
||
rich_diversity: float = Field(ge=0, le=2, description="表达多样性")
|
||
|
||
# 九、出版就绪度（小计最高 4；由原 5 收紧）
|
||
pub_editorial_cost: float = Field(ge=0, le=2, description="编辑成本")
|
||
pub_completeness: float = Field(ge=0, le=2, description="完整度")
|
||
|
||
total_score: float = Field(ge=0, le=100)
|
||
rationale: str = ""
|
||
|
||
major_strengths: list[str] = Field(default_factory=list)
|
||
major_issues: list[str] = Field(default_factory=list)
|
||
insufficient_evidence: list[str] = Field(default_factory=list)
|
||
evidence_refs: list[JudgeEvidenceRef] = Field(default_factory=list)
|
||
confidence: float = Field(default=0.75, ge=0.0, le=1.0)
|
||
|
||
authenticity_score: float = Field(default=0, ge=0, le=23)
|
||
information_score: float = Field(default=0, ge=0, le=14)
|
||
narrative_score: float = Field(default=0, ge=0, le=14)
|
||
language_score: float = Field(default=0, ge=0, le=18)
|
||
emotion_score: float = Field(default=0, ge=0, le=9)
|
||
character_score: float = Field(default=0, ge=0, le=9)
|
||
coherence_score: float = Field(default=0, ge=0, le=4)
|
||
richness_score: float = Field(default=0, ge=0, le=5)
|
||
publish_ready_score: float = Field(default=0, ge=0, le=4)
|
||
|
||
@model_validator(mode="before")
|
||
@classmethod
|
||
def _coerce_null_lists(cls, data: Any) -> Any:
|
||
if isinstance(data, dict):
|
||
for key in ("major_strengths", "major_issues", "insufficient_evidence"):
|
||
data[key] = _coerce_judge_str_list(data.get(key))
|
||
if data.get("evidence_refs") is None:
|
||
data["evidence_refs"] = []
|
||
return data
|
||
|
||
@model_validator(mode="after")
|
||
def _cap_meta_fields_and_sync_totals(self) -> Self:
|
||
def _cap_str_list(xs: list[str], *, max_items: int, max_chars: int) -> list[str]:
|
||
out: list[str] = []
|
||
for x in xs[:max_items]:
|
||
s = str(x).strip()
|
||
if s:
|
||
out.append(s[:max_chars])
|
||
return out
|
||
|
||
object.__setattr__(
|
||
self,
|
||
"major_strengths",
|
||
_cap_str_list(self.major_strengths, max_items=8, max_chars=200),
|
||
)
|
||
object.__setattr__(
|
||
self,
|
||
"major_issues",
|
||
_cap_str_list(self.major_issues, max_items=10, max_chars=200),
|
||
)
|
||
object.__setattr__(
|
||
self,
|
||
"insufficient_evidence",
|
||
_cap_str_list(self.insufficient_evidence, max_items=12, max_chars=200),
|
||
)
|
||
refs = list(self.evidence_refs)[:12]
|
||
object.__setattr__(self, "evidence_refs", refs)
|
||
|
||
authenticity = (
|
||
self.mem_fidelity
|
||
+ self.mem_factual_accuracy
|
||
+ self.mem_factual_coverage
|
||
+ self.mem_traceability
|
||
)
|
||
information = (
|
||
self.info_slot_coverage + self.info_sufficiency + self.info_density
|
||
)
|
||
narrative = self.narr_structure + self.narr_paragraphs + self.narr_pacing
|
||
language = (
|
||
self.lang_fluency
|
||
+ self.lang_conciseness
|
||
+ self.lang_literary
|
||
+ self.lang_controlled_expansion
|
||
+ self.lang_detail
|
||
+ self.lang_style
|
||
)
|
||
emotion = self.emo_authenticity + self.emo_depth
|
||
character = (
|
||
self.char_understanding + self.char_consistency + self.char_integration
|
||
)
|
||
coherence = self.coh_timeline + self.coh_cross_chapter
|
||
richness = self.rich_analogy + self.rich_diversity
|
||
publish = self.pub_editorial_cost + self.pub_completeness
|
||
expected = (
|
||
authenticity
|
||
+ information
|
||
+ narrative
|
||
+ language
|
||
+ emotion
|
||
+ character
|
||
+ coherence
|
||
+ richness
|
||
+ publish
|
||
)
|
||
synced = max(0.0, min(100.0, round(float(expected), 2)))
|
||
object.__setattr__(self, "total_score", synced)
|
||
object.__setattr__(self, "authenticity_score", authenticity)
|
||
object.__setattr__(self, "information_score", information)
|
||
object.__setattr__(self, "narrative_score", narrative)
|
||
object.__setattr__(self, "language_score", language)
|
||
object.__setattr__(self, "emotion_score", emotion)
|
||
object.__setattr__(self, "character_score", character)
|
||
object.__setattr__(self, "coherence_score", coherence)
|
||
object.__setattr__(self, "richness_score", richness)
|
||
object.__setattr__(self, "publish_ready_score", publish)
|
||
return self
|