309a051038
数据库与模型：新增多版迁移（章节证据快照、对话血缘、记忆事实/时间线 lineage 等），把「成稿 ↔ 对话/记忆」的溯源信息落到表结构里。 业务链路：会话与 WS、回忆录/故事流水线、记忆写入与 enrichment 等跟着接上线索与快照；新增章节证据快照与评测侧 EvalTraceService 等模块，方便组评审用的证据包。 内部评测：自动化 run 与手工 memoir 评审共用可追溯证据；rubric/ judge 相关脚本与文档有配套调整。 app-eval-web：Memoir/实验详情里能展开看证据摘要与 evidence_trace（含对话轮次 id）；Vite 代理与 development.sh 注入的 API 端口与当前默认内部评测端口一致，避免改端口后页面连错服务。 工程杂项：GitHub Actions / 仓库说明有更新；各适配器与支付/配额/plan 等多处为小改动或跟随主改动的收尾；新增/扩充了?
288 lines
8.9 KiB
Python
288 lines
8.9 KiB
Python
"""
|
||
证据包组装：跨 memory + story 的检索结果合并（业务层，非纯 repo）。
|
||
|
||
权威层级（可靠性 hardening）：
|
||
- **Chunk 原文**（未 excluded）为首要证据；rolling 摘要/故事摘录为便利视图，不得压过冲突的 chunk。
|
||
- **MemoryFact**：`confirmed` 为检索默认集；`candidate` 可被上游提升；`stale` 由 compaction 等标出，检索时应排除。
|
||
- 事实 ILIKE 无命中时是否退回「最近事实」由 `memory_fact_search_use_recent_fallback` 控制（默认可避免串台）。
|
||
|
||
Celery 使用 sync + 向量 chunks；`HybridRetriever` 使用 async + 向量 chunks。
|
||
"""
|
||
|
||
from __future__ import annotations
|
||
|
||
from typing import TYPE_CHECKING
|
||
|
||
from sqlalchemy.ext.asyncio import AsyncSession
|
||
from sqlalchemy.orm import Session
|
||
|
||
from app.core.config import settings
|
||
from app.core.logging import get_logger
|
||
from app.features.memory.repo import (
|
||
list_summaries_for_evidence_async,
|
||
list_summaries_for_evidence_sync,
|
||
search_chunks_vector_sync,
|
||
search_facts_for_user_async,
|
||
search_facts_for_user_sync,
|
||
search_timeline_events_for_user_async,
|
||
search_timeline_events_for_user_sync,
|
||
)
|
||
from app.features.story.repo import (
|
||
list_recent_stories_for_evidence,
|
||
list_recent_stories_for_evidence_sync,
|
||
)
|
||
|
||
if TYPE_CHECKING:
|
||
from app.ports.embedding import EmbeddingProvider
|
||
|
||
logger = get_logger(__name__)
|
||
|
||
EMPTY_EVIDENCE_BUNDLE: dict = {
|
||
"relevant_chunks": [],
|
||
"relevant_summaries": [],
|
||
"relevant_facts": [],
|
||
"timeline_hints": [],
|
||
"relevant_stories": [],
|
||
}
|
||
|
||
|
||
def _facts_to_dicts(facts) -> list[dict]:
|
||
return [
|
||
{
|
||
"id": f.id,
|
||
"fact_type": f.fact_type,
|
||
"subject": f.subject,
|
||
"predicate": f.predicate,
|
||
"object_json": f.object_json,
|
||
}
|
||
for f in facts
|
||
]
|
||
|
||
|
||
def _timeline_to_dicts(events) -> list[dict]:
|
||
return [
|
||
{
|
||
"id": e.id,
|
||
"event_year": e.event_year,
|
||
"event_date": e.event_date,
|
||
"title": e.title,
|
||
"description": e.description,
|
||
}
|
||
for e in events
|
||
]
|
||
|
||
|
||
def _stories_to_dicts(story_rows) -> list[dict]:
|
||
return [
|
||
{
|
||
"id": s.id,
|
||
"title": s.title,
|
||
"summary": s.summary,
|
||
"stage": s.stage,
|
||
"story_type": s.story_type,
|
||
}
|
||
for s in story_rows
|
||
]
|
||
|
||
|
||
def fetch_evidence_metadata_sync(
|
||
session: Session, user_id: str, q: str, top_k: int
|
||
) -> dict:
|
||
"""非 chunk 证据：摘要、事实、时间线、故事（sync）。"""
|
||
facts = search_facts_for_user_sync(session, user_id, q, top_k)
|
||
events = search_timeline_events_for_user_sync(session, user_id, q, top_k)
|
||
relevant_summaries = list_summaries_for_evidence_sync(
|
||
session, user_id=user_id, q=q, limit=top_k
|
||
)
|
||
story_rows = list_recent_stories_for_evidence_sync(
|
||
session, user_id, query=q, limit=top_k
|
||
)
|
||
return {
|
||
"relevant_facts": _facts_to_dicts(facts),
|
||
"timeline_hints": _timeline_to_dicts(events),
|
||
"relevant_summaries": relevant_summaries,
|
||
"relevant_stories": _stories_to_dicts(story_rows),
|
||
}
|
||
|
||
|
||
async def fetch_evidence_metadata_async(
|
||
db: AsyncSession, user_id: str, q: str, top_k: int
|
||
) -> dict:
|
||
"""非 chunk 证据（async）。"""
|
||
facts = await search_facts_for_user_async(db, user_id, q, top_k)
|
||
events = await search_timeline_events_for_user_async(db, user_id, q, top_k)
|
||
relevant_summaries = await list_summaries_for_evidence_async(
|
||
db, user_id=user_id, q=q, limit=top_k
|
||
)
|
||
story_rows = await list_recent_stories_for_evidence(
|
||
db, user_id=user_id, query=q, limit=top_k
|
||
)
|
||
return {
|
||
"relevant_facts": _facts_to_dicts(facts),
|
||
"timeline_hints": _timeline_to_dicts(events),
|
||
"relevant_summaries": relevant_summaries,
|
||
"relevant_stories": _stories_to_dicts(story_rows),
|
||
}
|
||
|
||
|
||
def _empty_query_bundle_sync(session: Session, user_id: str, top_k: int) -> dict:
|
||
"""空 query 时的「浏览」降级：rolling 摘要 + 事实/时间线 fallback。"""
|
||
from sqlalchemy import select
|
||
|
||
from app.features.memory.models import MemorySummary
|
||
from app.features.memory.repo import (
|
||
get_facts_for_user_sync,
|
||
get_timeline_events_for_user_sync,
|
||
)
|
||
|
||
rolling = (
|
||
session.execute(
|
||
select(MemorySummary)
|
||
.where(
|
||
MemorySummary.user_id == user_id,
|
||
MemorySummary.summary_type == "rolling",
|
||
)
|
||
.order_by(MemorySummary.updated_at.desc())
|
||
.limit(1)
|
||
)
|
||
.unique()
|
||
.scalar_one_or_none()
|
||
)
|
||
summaries = []
|
||
if rolling:
|
||
summaries = [
|
||
{
|
||
"id": rolling.id,
|
||
"summary_type": rolling.summary_type,
|
||
"content": rolling.content,
|
||
"source_chunk_ids": rolling.source_chunk_ids,
|
||
}
|
||
]
|
||
facts = get_facts_for_user_sync(session, user_id, top_k)
|
||
events = get_timeline_events_for_user_sync(session, user_id, top_k)
|
||
return {
|
||
"relevant_chunks": [],
|
||
"relevant_summaries": summaries,
|
||
"relevant_facts": _facts_to_dicts(facts),
|
||
"timeline_hints": _timeline_to_dicts(events),
|
||
"relevant_stories": [],
|
||
}
|
||
|
||
|
||
async def _empty_query_bundle_async(db: AsyncSession, user_id: str, top_k: int) -> dict:
|
||
from sqlalchemy import select
|
||
|
||
from app.features.memory.models import MemorySummary
|
||
from app.features.memory.repo import (
|
||
get_facts_for_user,
|
||
get_timeline_events_for_user,
|
||
)
|
||
|
||
roll_stmt = (
|
||
select(MemorySummary)
|
||
.where(
|
||
MemorySummary.user_id == user_id,
|
||
MemorySummary.summary_type == "rolling",
|
||
)
|
||
.order_by(MemorySummary.updated_at.desc())
|
||
.limit(1)
|
||
)
|
||
r_result = await db.execute(roll_stmt)
|
||
rolling = r_result.unique().scalar_one_or_none()
|
||
summaries = []
|
||
if rolling:
|
||
summaries = [
|
||
{
|
||
"id": rolling.id,
|
||
"summary_type": rolling.summary_type,
|
||
"content": rolling.content,
|
||
"source_chunk_ids": rolling.source_chunk_ids,
|
||
}
|
||
]
|
||
facts = await get_facts_for_user(db, user_id=user_id, limit=top_k)
|
||
events = await get_timeline_events_for_user(db, user_id=user_id, limit=top_k)
|
||
return {
|
||
"relevant_chunks": [],
|
||
"relevant_summaries": summaries,
|
||
"relevant_facts": _facts_to_dicts(facts),
|
||
"timeline_hints": _timeline_to_dicts(events),
|
||
"relevant_stories": [],
|
||
}
|
||
|
||
|
||
def retrieve_evidence_bundle_sync(
|
||
session: Session,
|
||
user_id: str,
|
||
query: str,
|
||
*,
|
||
top_k: int = 10,
|
||
embedding_provider: "EmbeddingProvider | None" = None,
|
||
) -> dict:
|
||
"""Celery / 叙事流水线：向量 chunks + 元数据（需 embedding_provider）。"""
|
||
if not query or not query.strip():
|
||
if settings.memory_evidence_empty_query_include_rolling:
|
||
return _empty_query_bundle_sync(session, user_id, top_k)
|
||
return dict(EMPTY_EVIDENCE_BUNDLE)
|
||
q = query.strip()
|
||
relevant_chunks: list[dict] = []
|
||
if embedding_provider is not None:
|
||
try:
|
||
q_emb = embedding_provider.embed_text_sync(q)
|
||
except Exception as exc:
|
||
logger.warning(
|
||
"retrieve_evidence_bundle_sync embed failed user_id={} err={}",
|
||
user_id,
|
||
exc,
|
||
)
|
||
q_emb = []
|
||
if q_emb:
|
||
chunk_rows = search_chunks_vector_sync(session, user_id, q_emb, top_k)
|
||
relevant_chunks = [
|
||
{
|
||
"id": r["id"],
|
||
"content": r["content"],
|
||
"chunk_index": r["chunk_index"],
|
||
}
|
||
for r in chunk_rows
|
||
]
|
||
else:
|
||
logger.warning(
|
||
"retrieve_evidence_bundle_sync empty_query_embedding user_id={}",
|
||
user_id,
|
||
)
|
||
else:
|
||
logger.warning(
|
||
"retrieve_evidence_bundle_sync no_embedding_provider user_id={}",
|
||
user_id,
|
||
)
|
||
meta = fetch_evidence_metadata_sync(session, user_id, q, top_k)
|
||
return {
|
||
"relevant_chunks": relevant_chunks,
|
||
**meta,
|
||
}
|
||
|
||
|
||
async def retrieve_evidence_bundle_async(
|
||
db: AsyncSession,
|
||
user_id: str,
|
||
query: str,
|
||
*,
|
||
top_k: int = 10,
|
||
merged_chunk_dicts: list[dict],
|
||
) -> dict:
|
||
"""
|
||
异步路径：chunk 已由调用方（如 HybridRetriever）向量检索填入；此处只拼元数据。
|
||
|
||
merged_chunk_dicts: [{"id","content","chunk_index"}, ...]
|
||
"""
|
||
if not query or not query.strip():
|
||
if settings.memory_evidence_empty_query_include_rolling:
|
||
return await _empty_query_bundle_async(db, user_id, top_k)
|
||
return dict(EMPTY_EVIDENCE_BUNDLE)
|
||
q = query.strip()
|
||
meta = await fetch_evidence_metadata_async(db, user_id, q, top_k)
|
||
return {
|
||
"relevant_chunks": merged_chunk_dicts,
|
||
**meta,
|
||
}
|