kevin/life-echo
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(api): 回忆录管线简化、路由延迟池与相关加固

Browse Source 
- Phase1/2：移除 MemoirOrchestrator.run 与 process_memoir_segments 别名；文档改为 process_memoir_phase1。
- 槽位校验集中到 stage_constants（filter_stage_slots），批处理与顺序路径及 state_service 写库一致。
- StoryRoute：no_llm/parse_error/invalid_target 保守 new_story；短篇护栏不覆盖这些 fallback。
- Phase2 低置信单路径可选延迟（StoryPipelineResult.deferred）：不写 Chapter/Story，Segment 记录 defer 元数据，冷却内不重复消费；上限后停自动重试，Phase1 同类目新段唤醒池内段。
- Alembic 0017：segments 表 narrative_defer_* 列。
- ProfileAgent：经 LlmGateway/注入 Provider 统一聊天与 JSON，新增测试。
- ImagePromptOrchestrator：LLM 初始化失败可依配置降级或硬失败；补充策略测试。
- 配套单测与 README/本地开发文档表述更新。

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
Kevin
2026-05-06 13:18:02 +08:00
parent 3234396254
commit 59d4b19d7d
24 changed files with 1182 additions and 183 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
api/app/agents/memoir/batch_phase1_prep.py
View File

@@ -10,7 +10,6 @@ from typing import Any, Callable, Dict, List
from app.agents.memoir.prompts import get_batch_memoir_phase1_prep_prompt
from app.agents.memoir.schemas import BatchPhase1LLMOutput
from app.agents.stage_constants import STAGE_SLOT_KEYS
from app.agents.state_schema import MemoirStateSchema
from app.core.config import settings
from app.core.llm_call import LLMCallError, llm_json_call
@@ -19,11 +18,6 @@ from app.features.conversation.models import Segment
logger = get_logger(__name__)
STAGE_ALLOWED_SLOTS: Dict[str, frozenset[str]] = {
k: frozenset(v) for k, v in STAGE_SLOT_KEYS.items()
}
def _slots_snapshot(state: MemoirStateSchema) -> dict:
snap: dict = {}
for stage, buckets in (state.slots or {}).items():

101
api/app/agents/memoir/orchestrator.py
View File

@@ -8,12 +8,9 @@ from __future__ import annotations
import time
from dataclasses import dataclass
from typing import Any, Callable, Dict, List, Optional, Set, Tuple
from typing import Any, Callable, Dict, List, Optional, Set
from app.agents.memoir.batch_phase1_prep import (
STAGE_ALLOWED_SLOTS,
run_batch_phase1_prep_chunked,
)
from app.agents.memoir.batch_phase1_prep import run_batch_phase1_prep_chunked
from app.agents.memoir.classification_agent import (
ClassificationAgent,
_looks_like_fragment_only,
@@ -22,7 +19,11 @@ from app.agents.memoir.classification_agent import (
_detect_stage as detect_stage_from_keywords,
)
from app.agents.memoir.extraction_agent import ExtractionAgent, ExtractionResult
from app.agents.stage_constants import normalize_chapter_category, normalize_chat_stage
from app.agents.stage_constants import (
filter_stage_slots,
normalize_chapter_category,
normalize_chat_stage,
)
from app.agents.state_schema import MemoirStateSchema
from app.core.agent_logging import agent_span, agent_summary_enabled, log_agent_detail
from app.core.config import settings
@@ -92,7 +93,7 @@ class MemoirOrchestrator:
)
if use_batch:
try:
result = self._prepare_batches_via_batch_llm(
prepared_batch = self._prepare_batches_via_batch_llm(
segments=segments,
state=state,
classify_extract_llm=classify_extract_llm,
@@ -104,7 +105,7 @@ class MemoirOrchestrator:
"msg=Phase1 批处理 LLM 路径已使用",
len(segments),
)
return result
return prepared_batch
except Exception as e:
logger.warning(
"event=phase1_batch_path_fallback segment_count={} exc={} "
@@ -132,8 +133,12 @@ class MemoirOrchestrator:
stage_slots=stage_slots_raw,
llm=classify_extract_llm,
)
detected_stage = result.detected_stage
for slot_name, snippet in result.slots.items():
fb = state.current_stage or "childhood"
detected_stage = normalize_chat_stage(result.detected_stage, fb)
result_slots = filter_stage_slots(detected_stage, result.slots, fb)
if not result_slots:
detected_stage = normalize_chat_stage(fb, fb)
for slot_name, snippet in result_slots.items():
state = update_slot(detected_stage, slot_name, snippet, [segment.id])
with agent_span(
@@ -148,7 +153,7 @@ class MemoirOrchestrator:
segment_id=segment.id,
)
chapter_category = classify_result.category
if (not result.slots) and classify_result.llm_said_none:
if (not result_slots) and classify_result.llm_said_none:
segment_skip_story_ids.add(str(segment.id))
segment_chapter_category[str(segment.id)] = chapter_category
@@ -166,7 +171,7 @@ class MemoirOrchestrator:
logger,
"MemoirOrchestrator.segment_done segment_id={} slots={}",
segment.id,
list((result.slots or {}).keys()),
list(result_slots.keys()),
)
category_to_segments.setdefault(chapter_category, []).append(segment)
@@ -211,8 +216,7 @@ class MemoirOrchestrator:
else:
detected_stage = normalize_chat_stage(row.detected_stage, fb)
allowed = STAGE_ALLOWED_SLOTS.get(detected_stage, frozenset())
result_slots = {k: v for k, v in result_slots.items() if k in allowed}
result_slots = filter_stage_slots(detected_stage, result_slots, fb)
if not result_slots:
detected_stage = normalize_chat_stage(fb, fb)
@@ -269,72 +273,3 @@ class MemoirOrchestrator:
segment_skip_story_ids=segment_skip_story_ids,
segment_chapter_category=segment_chapter_category,
)
def run(
self,
*,
segments: List[Segment],
llm: Any,
user_profile: str = "",
user_birth_year: Any = None,
get_or_create_state: Callable[[], MemoirStateSchema],
update_slot: Callable[[str, str, str, List[str]], MemoirStateSchema],
acquire_lock: Callable[[str], bool],
release_lock: Callable[[str], None],
process_category: Callable[
[
str,
List[Segment],
MemoirStateSchema,
str,
Any,
Any,
],
Tuple[Any, bool],
],
raise_retry: Callable[[], None],
llm_fast: Any | None = None,
) -> Tuple[Set[str], int]:
"""
执行回忆录流水线。
process_category(category, segments, state, user_profile, user_birth_year, llm)
返回 (chapter, has_images_to_generate)。
返回 (chapters_to_enqueue, processed_count)。
raise_retry 用于锁竞争时抛出 Celery retry。
"""
prepared = self.prepare_batches(
segments=segments,
llm=llm,
llm_fast=llm_fast,
get_or_create_state=get_or_create_state,
update_slot=update_slot,
on_phase1_chunk=None,
)
state = prepared.state
chapters_to_enqueue: Set[str] = set()
category_to_segments = prepared.category_to_segments
# 按 category 调用 process_category：叙事生成、持久化、封面入队标记
for chapter_category, category_segments in category_to_segments.items():
if not acquire_lock(chapter_category):
logger.warning(
"章节锁竞争: category={}, 延迟重试",
chapter_category,
)
raise_retry()
try:
chapter, has_images = process_category(
chapter_category,
category_segments,
state,
user_profile,
user_birth_year,
llm,
)
if chapter and has_images:
chapters_to_enqueue.add(chapter.id)
finally:
release_lock(chapter_category)
return chapters_to_enqueue, len(segments)

33
api/app/agents/memoir/story_route_agent.py
View File

@@ -31,6 +31,9 @@ PLAN_BATCH_MAX_SEGMENTS = 48
# 童年 / 求学 / 家庭：模型与后处理均倾向「少拆分、优先续写」
APPEND_FIRST_CHAPTER_CATEGORIES = frozenset({"childhood", "education", "family"})
# These route outcomes are conservative fail-safes, not semantic append matches.
FALLBACK_NEW_STORY_REASONS = frozenset({"no_llm", "parse_error", "invalid_target"})
def default_append_target_story_id(
candidate_stories: list[Story],
@@ -220,13 +223,6 @@ class StoryRouteAgent:
story_meta: dict[str, dict[str, int]] | None = None,
) -> StoryRouteDecision:
if not llm:
fb = default_append_target_story_id(candidate_stories, story_meta, settings)
if fb and fb in valid_story_ids:
return StoryRouteDecision(
decision="append_story",
target_story_id=fb,
reason="no_llm_default_append",
)
return StoryRouteDecision(
decision="new_story",
new_story_title=None,
@@ -241,13 +237,6 @@ class StoryRouteAgent:
)
def _decide_fallback() -> StoryRouteDecision:
fb = default_append_target_story_id(candidate_stories, story_meta, settings)
if fb and fb in valid_story_ids:
return StoryRouteDecision(
decision="append_story",
target_story_id=fb,
reason="parse_error_default_append",
)
return StoryRouteDecision(
decision="new_story",
new_story_title=None,
@@ -266,22 +255,8 @@ class StoryRouteAgent:
if decision.decision == "append_story":
tid = decision.target_story_id
if not tid or tid not in valid_story_ids:
fb = default_append_target_story_id(
candidate_stories, story_meta, settings
)
if fb and fb in valid_story_ids:
logger.info(
"StoryRoute append 无效 target_story_id={}，回退默认 append {}",
tid,
fb,
)
return StoryRouteDecision(
decision="append_story",
target_story_id=fb,
reason="invalid_target_default_append",
)
logger.warning(
"StoryRoute append 无效 target_story_id={}，且无可用默认目标，回退 new_story",
"StoryRoute append 无效 target_story_id={}，回退 new_story",
tid,
)
return StoryRouteDecision(