89 lines
3.0 KiB
Python
89 lines
3.0 KiB
Python
|
"""
|
|||
|
|
FidelityCheckAgent：比较「用户口述」与叙事 JSON 输出，判定是否存在明显编造或越界。
|
|||
|
|
失败时由流水线回退为口述正文（见 story_pipeline_sync）。
|
|||
|
|
"""
|
|||
|
|
|
|||
|
|
from __future__ import annotations
|
|||
|
|
|
|||
|
|
import json
|
|||
|
|
import re
|
|||
|
|
from typing import Any
|
|||
|
|
|
|||
|
|
from app.core.config import settings
|
|||
|
|
from app.core.langchain_llm import invoke_json_object
|
|||
|
|
from app.core.logging import get_logger
|
|||
|
|
from app.features.memoir.memoir_images.json_payload import extract_json_payload
|
|||
|
|
|
|||
|
|
logger = get_logger(__name__)
|
|||
|
|
|
|||
|
|
# 生成稿中出现的四位年份，若口述中未出现同串，仅打日志（不误杀）
|
|||
|
|
_YEAR_4_RE = re.compile(r"(?<!\d)(19|20)\d{2}(?!\d)")
|
|||
|
|
|
|||
|
|
|
|||
|
|
def _log_suspicious_years_not_in_oral(oral_text: str, narrative_json: str) -> None:
|
|||
|
|
oral = oral_text or ""
|
|||
|
|
gen = narrative_json or ""
|
|||
|
|
for m in _YEAR_4_RE.finditer(gen):
|
|||
|
|
y = m.group(0)
|
|||
|
|
if y not in oral:
|
|||
|
|
logger.debug(
|
|||
|
|
"event=fidelity_heuristic_year_not_in_oral year={} oral_len={} gen_len={}",
|
|||
|
|
y,
|
|||
|
|
len(oral),
|
|||
|
|
len(gen),
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
|
|||
|
|
class FidelityCheckAgent:
|
|||
|
|
"""叙事忠实度检查（json_object）；失败时上层应回退为口述原文。"""
|
|||
|
|
|
|||
|
|
def passes(
|
|||
|
|
self,
|
|||
|
|
*,
|
|||
|
|
oral_text: str,
|
|||
|
|
narrative_json: str,
|
|||
|
|
llm: Any,
|
|||
|
|
) -> bool:
|
|||
|
|
if not llm or not settings.memoir_fidelity_check_enabled:
|
|||
|
|
return True
|
|||
|
|
oral = (oral_text or "").strip()
|
|||
|
|
gen = (narrative_json or "").strip()
|
|||
|
|
if not oral or not gen:
|
|||
|
|
return True
|
|||
|
|
_log_suspicious_years_not_in_oral(oral, gen)
|
|||
|
|
prompt = f"""你是事实核对员。比较下面两段文字。
|
|||
|
|
|
|||
|
|
【用户口述】（亲历内容）
|
|||
|
|
{oral[:8000]}
|
|||
|
|
|
|||
|
|
【模型生成的 JSON 叙事】（应只含口述中已有事实的整理，不得添油加醋）
|
|||
|
|
{gen[:16000]}
|
|||
|
|
|
|||
|
|
判断：生成稿是否出现**口述中明显没有**的具体人名、地名、时间、数字、事件经过、对话，或把摘录/档案里才有的信息写成了用户亲口经历？
|
|||
|
|
若存在明显编造或越界，pass=false；若仅口语转书面、删赘词、合并指代，pass=true。
|
|||
|
|
|
|||
|
|
**JSON 输出**：只输出一个合法 JSON 对象。
|
|||
|
|
{{"pass": true, "reason": null}}
|
|||
|
|
或
|
|||
|
|
{{"pass": false, "reason": "一句话说明"}}
|
|||
|
|
|
|||
|
|
只输出 JSON，不要其它文字。"""
|
|||
|
|
try:
|
|||
|
|
raw = invoke_json_object(
|
|||
|
|
llm,
|
|||
|
|
prompt,
|
|||
|
|
max_tokens=settings.memoir_fidelity_check_max_tokens,
|
|||
|
|
agent="FidelityCheckAgent.passes",
|
|||
|
|
)
|
|||
|
|
data = json.loads(extract_json_payload(raw))
|
|||
|
|
ok = bool(data.get("pass", True))
|
|||
|
|
if not ok:
|
|||
|
|
logger.warning(
|
|||
|
|
"event=fidelity_check_fail reason={}",
|
|||
|
|
(data.get("reason") or "")[:200],
|
|||
|
|
)
|
|||
|
|
return ok
|
|||
|
|
except Exception as e:
|
|||
|
|
logger.warning("FidelityCheckAgent 解析失败，放行: {}", e)
|
|||
|
|
return True
|