feat(api+app): 对话阶段化、回忆录流水线与客户端会话体验

- DB: segments 用户输入文本(Alembic 0002)
- Chat: 阶段检测/阶段提示/回复限制,编排与访谈/画像 prompts 调整
- Memoir: 忠实度检查 agent,叙事与分类等链路更新
- Core: agent 日志、Alembic 启动、LangChain/日志/配置等
- Story: time_hints;Memory 检索与相关测试
- Expo: 助手头像、会话页与消息拆分、实时会话与文案/i18n
- Docs/scripts/tests: 迁移脚本、LLM JSON/记忆检索文档、新增单测
This commit is contained in:
Kevin
2026-03-26 12:13:36 +08:00
parent 49b089354c
commit a3f61fcc0f
94 changed files with 3332 additions and 672 deletions

View File

@@ -0,0 +1,88 @@
"""
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}}
{{"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