Merge branch 'eval/elapsed-time-memoir-batch-chunk' into development

api/app/features/memoir/story_pipeline_sync.py

@@ -10,9 +10,10 @@ import json
 import re
 import time
 import uuid
+from concurrent.futures import ThreadPoolExecutor
 from typing import Any
 
-from sqlalchemy import select
+from sqlalchemy import func, select
 from sqlalchemy.orm import Session, joinedload
 
 from app.agents.memoir.narrative_agent import NarrativeAgent
@@ -405,11 +406,13 @@ def _gate_narrative_fidelity(
     llm: Any,
     *,
     existing_canonical: str | None = None,
+    fidelity_llm: Any | None = None,
 ) -> tuple[str, str]:
     """返回 (文本, fallback 原因)；忠实度不通过时第二项为 fidelity_failed。"""
     from app.agents.memoir.fidelity_check_agent import FidelityCheckAgent
 
-    if not settings.memoir_fidelity_check_enabled or not llm:
+    check_llm = fidelity_llm if fidelity_llm is not None else llm
+    if not settings.memoir_fidelity_check_enabled or not check_llm:
         return narrative_raw, "none"
     agent = FidelityCheckAgent()
     ex = (existing_canonical or "").strip() or None
@@ -417,7 +420,7 @@ def _gate_narrative_fidelity(
     if agent.passes(
         oral_text=oral_text,
         narrative_json=narrative_raw,
-        llm=llm,
+        llm=check_llm,
         existing_canonical_markdown=ex,
         is_append=is_append,
     ):
@@ -564,14 +567,23 @@ def _merge_fallback_type(gate_ft: str, apply_ft: str) -> str:
 def _story_meta_for_route(
     session: Session, candidates: list
 ) -> dict[str, dict[str, int]]:
-    meta: dict[str, dict[str, int]] = {}
-    for s in candidates:
-        sid = str(s.id)
-        meta[sid] = {
+    if not candidates:
+        return {}
+    sids = [str(s.id) for s in candidates]
+    stmt = (
+        select(StoryVersion.story_id, func.count(StoryVersion.id))
+        .where(StoryVersion.story_id.in_(sids))
+        .group_by(StoryVersion.story_id)
+    )
+    rows = session.execute(stmt).all()
+    counts: dict[str, int] = {str(r[0]): int(r[1] or 0) for r in rows}
+    return {
+        str(s.id): {
             "char_count": len((s.canonical_markdown or "").strip()),
-            "version_count": count_story_versions_sync(session, sid),
+            "version_count": counts.get(str(s.id), 0),
         }
-    return meta
+        for s in candidates
+    }
 
 
 def _ensure_chapter_record(
@@ -615,7 +627,6 @@ def _ensure_chapter_record(
         )
         chapter.is_new = True
         session.flush()
-    refresh_chapter_evidence_snapshot_with_retry_sync(session, str(chapter.id))
     return chapter
 
 
@@ -710,6 +721,7 @@ def _execute_narrative_unit(
     background_voice: str = "default",
     occupation: str = "",
     memoir_correlation_id: str | None = None,
+    fidelity_llm: Any | None = None,
 ) -> tuple[str | None, bool]:
     """
     Unified narrative unit executor: generate narrative, apply fidelity/safety,
@@ -744,6 +756,7 @@ def _execute_narrative_unit(
         raw_gen,
         llm,
         existing_canonical=existing_for_narrative or None,
+        fidelity_llm=fidelity_llm,
     )
     narrative_raw, fb_apply = _apply_narrative_fallbacks(
         narrative_raw,
@@ -792,16 +805,7 @@ def _execute_narrative_unit(
         sid_log = target_story_id
         is_append = True
     else:
-        story_title = _maybe_generate_title(
-            narrative_agent,
-            chapter_category=chapter_category,
-            md=md,
-            slot_snippets=slot_snippets,
-            user_profile=user_profile,
-            user_birth_year=user_birth_year,
-            llm=llm,
-            oral_scope=oral_norm,
-        )
+        story_title = _placeholder_title(chapter_category)
         st = create_story_with_version_sync(
             session,
             user_id=user_id,
@@ -809,6 +813,21 @@ def _execute_narrative_unit(
             canonical_markdown=md,
             stage=chapter_category,
         )
+        try:
+            from app.tasks.story_title_tasks import generate_story_title_after_create
+
+            generate_story_title_after_create.delay(
+                str(st.id),
+                chapter_category,
+                oral_norm,
+                user_id,
+            )
+        except Exception as exc:
+            logger.warning(
+                "event=story_title_enqueue_failed story_id={} err={}",
+                st.id,
+                exc,
+            )
         ensure_chapter_story_link_sync(
             session, chapter_id=str(chapter.id), story_id=str(st.id)
         )
@@ -874,6 +893,7 @@ def _run_batch_plan_writes(
     background_voice: str = "default",
     occupation: str = "",
     memoir_correlation_id: str | None = None,
+    fidelity_llm: Any | None = None,
 ) -> set[str]:
     dispatch_ids: set[str] = set()
     for unit in plan.units:
@@ -919,6 +939,7 @@ def _run_batch_plan_writes(
             background_voice=background_voice,
             occupation=occupation,
             memoir_correlation_id=memoir_correlation_id,
+            fidelity_llm=fidelity_llm,
         )
         if sid:
             dispatch_ids.add(sid)
@@ -938,6 +959,7 @@ def run_story_pipeline_for_category_batch(
     background_voice: str = "default",
     occupation: str = "",
     memoir_correlation_id: str | None = None,
+    llm_fast: Any | None = None,
 ) -> tuple[Chapter | None, bool, set[str]]:
     """
     返回 (chapter, needs_cover_enqueue, story_ids_to_dispatch_after_commit)。
@@ -946,6 +968,8 @@ def run_story_pipeline_for_category_batch(
     narrative_agent = NarrativeAgent()
     route_agent = StoryRouteAgent()
     dispatch_ids: set[str] = set()
+    llm_route = llm_fast if llm_fast is not None else llm
+    llm_fidelity = llm_fast if llm_fast is not None else llm
 
     segment_texts = [seg.user_input_text or "" for seg in category_segments]
     combined_text = "\n\n".join(segment_texts)
@@ -957,25 +981,40 @@ def run_story_pipeline_for_category_batch(
         top_k = int(settings.evidence_top_k_large_batch)
     emb = get_embedding_provider()
     embedding_available = emb.is_available()
-    _t0 = time.perf_counter()
-    try:
-        evidence = retrieve_evidence_sync(
-            session,
-            user_id,
-            combined_text,
-            top_k=top_k,
-            embedding_provider=emb,
-        )
-    except Exception as e:
-        logger.warning("Evidence 检索跳过: {}", e)
-        evidence = {
-            "relevant_chunks": [],
-            "relevant_summaries": [],
-            "relevant_facts": [],
-            "timeline_hints": [],
-            "relevant_stories": [],
-        }
-    pipeline_phase_timings["evidence"] = time.perf_counter() - _t0
+
+    def _oral_job() -> tuple[str, float]:
+        t_oral = time.perf_counter()
+        out = normalize_oral_for_memoir(combined_text, llm=llm)
+        return out, time.perf_counter() - t_oral
+
+    _t_parallel = time.perf_counter()
+    with ThreadPoolExecutor(max_workers=1) as pool:
+        oral_future = pool.submit(_oral_job)
+        _t_ev = time.perf_counter()
+        try:
+            evidence = retrieve_evidence_sync(
+                session,
+                user_id,
+                combined_text,
+                top_k=top_k,
+                embedding_provider=emb,
+            )
+        except Exception as e:
+            logger.warning("Evidence 检索跳过: {}", e)
+            evidence = {
+                "relevant_chunks": [],
+                "relevant_summaries": [],
+                "relevant_facts": [],
+                "timeline_hints": [],
+                "relevant_stories": [],
+            }
+        ev_elapsed = time.perf_counter() - _t_ev
+        oral_for_memoir, oral_elapsed = oral_future.result()
+    pipeline_phase_timings["evidence"] = ev_elapsed
+    pipeline_phase_timings["oral_normalize"] = oral_elapsed
+    pipeline_phase_timings["evidence_oral_parallel_wall"] = (
+        time.perf_counter() - _t_parallel
+    )
 
     logger.info(
         "memoir_evidence_retrieved user_id={} chunks={} facts={} summaries={} stories={} vector_ok={}",
@@ -988,9 +1027,6 @@ def run_story_pipeline_for_category_batch(
     )
 
     evidence_text = format_evidence_chunks_for_prompt(evidence)
-    _t0 = time.perf_counter()
-    oral_for_memoir = normalize_oral_for_memoir(combined_text, llm=llm)
-    pipeline_phase_timings["oral_normalize"] = time.perf_counter() - _t0
     ct_raw = (combined_text or "").strip()
     om_norm = (oral_for_memoir or "").strip()
     if ct_raw != om_norm:
@@ -999,7 +1035,6 @@ def run_story_pipeline_for_category_batch(
             len(ct_raw),
             len(om_norm),
         )
-    new_content_input = format_narrative_user_content(oral_for_memoir, evidence_text)
     logger.info(
         "event=memoir_story_pipeline_start memoir_correlation_id={} user_id={} "
         "chapter_category={} segment_count={}",
@@ -1044,7 +1079,7 @@ def run_story_pipeline_for_category_batch(
 
     _t0 = time.perf_counter()
     use_batch_plan = (
-        llm
+        llm_route
         and len(category_segments) >= 2
         and len(category_segments) <= PLAN_BATCH_MAX_SEGMENTS
     )
@@ -1056,7 +1091,7 @@ def run_story_pipeline_for_category_batch(
             chapter_title=title,
             segments=segs,
             candidate_stories=candidates,
-            llm=llm,
+            llm=llm_route,
             valid_story_ids=valid_ids,
             story_meta=story_meta,
         )
@@ -1093,6 +1128,7 @@ def run_story_pipeline_for_category_batch(
             background_voice=background_voice,
             occupation=occupation,
             memoir_correlation_id=memoir_correlation_id,
+            fidelity_llm=llm_fidelity,
         )
     else:
         route = route_agent.decide(
@@ -1100,7 +1136,7 @@ def run_story_pipeline_for_category_batch(
             chapter_title=title,
             batch_transcript=route_transcript,
             candidate_stories=candidates,
-            llm=llm,
+            llm=llm_route,
             valid_story_ids=valid_ids,
             story_meta=story_meta,
         )
@@ -1145,6 +1181,7 @@ def run_story_pipeline_for_category_batch(
             background_voice=background_voice,
             occupation=occupation,
             memoir_correlation_id=memoir_correlation_id,
+            fidelity_llm=llm_fidelity,
         )
         if sid:
             dispatch_ids.add(sid)