"""组装 Chapter/Story 评测证据闭包并格式化为评审输入。""" from __future__ import annotations from typing import Literal from sqlalchemy.ext.asyncio import AsyncSession from app.features.conversation import repo as conversation_repo from app.features.conversation.lineage_schemas import aggregate_lineage_from_segments from app.features.evaluation.eval_trace_format import ( build_segment_transcript, format_chapter_for_judge, format_story_for_judge, ) from app.features.evaluation.eval_trace_repo import ( fetch_ai_messages_for_segments, fetch_memory_closure_for_conversations, fetch_segments_for_user, get_chapter_for_eval_trace, get_story_for_eval_trace, list_chapter_ids_for_story, load_chunks_by_ids, load_facts_by_ids, load_summaries_by_ids, load_timeline_by_ids, normalize_source_segment_ids, story_link_ids_by_type, ) from app.features.evaluation.eval_trace_schemas import ( ChapterEvidenceBundle, FormattedMemoirEvidence, StoryEvidenceBundle, ) from app.features.memoir.chapter_evidence_snapshot import ( EVIDENCE_SNAPSHOT_SCHEMA_VERSION, ) from app.features.memoir.models import Chapter from app.features.story.models import Story, StoryVersion _MAX_EVIDENCE_CONVERSATIONS = 8 _MAX_EVIDENCE_TRANSCRIPT_CHARS = 16_000 def _segments_in_order(segments: list, segment_ids: list[str]) -> list: order = {str(sid): i for i, sid in enumerate(segment_ids)} return sorted(segments, key=lambda s: order.get(str(s.id), 9999)) def _trim_fallback_transcript(text: str) -> str: s = (text or "").strip() if len(s) <= _MAX_EVIDENCE_TRANSCRIPT_CHARS: return s return f"{s[:_MAX_EVIDENCE_TRANSCRIPT_CHARS]}\n\n…（访谈证据已截断）" async def fallback_user_transcript_evidence(db: AsyncSession, user_id: str) -> str: """legacy：最近若干会话全文（仅作 fallback，调用方须声明 tier=fallback）。""" conversations = await conversation_repo.get_user_conversations(user_id, db) if not conversations: return "" parts: list[str] = [] for conv in reversed(conversations[:_MAX_EVIDENCE_CONVERSATIONS]): rows = await conversation_repo.get_conversation_messages(str(conv.id), db) blocks: list[str] = [] for row in rows: role = str(row.role or "").lower() body = (row.content or "").strip() if not body: continue label = "用户" if role == "human" else "AI" blocks.append(f"{label}: {body}") transcript = "\n\n".join(blocks) if transcript: parts.append(f"## 会话 {str(conv.id)}\n{transcript}") return _trim_fallback_transcript("\n\n".join(parts)) class EvalTraceService: def __init__(self, db: AsyncSession) -> None: self._db = db async def _story_dialogue_lineage( self, st: Story, segments: list, segment_ids_ordered: list[str], ) -> dict | None: if getattr(st, "current_version_id", None): ver = await self._db.get(StoryVersion, st.current_version_id) if ver and isinstance(getattr(ver, "lineage_json", None), dict): lj = ver.lineage_json if lj.get("turns"): return lj if segments and segment_ids_ordered: ordered = _segments_in_order(segments, segment_ids_ordered) conv_ids = sorted( {str(s.conversation_id) for s in segments if s.conversation_id} ) return aggregate_lineage_from_segments( ordered, conversation_id_fallback=conv_ids[0] if conv_ids else None ) return None def _chapter_closure_tier( self, *, segment_ids_resolved: list[str], chunk_ids: list[str], fact_ids: list[str], tl_ids: list[str], sum_ids: list[str], ) -> Literal["strict", "partial", "fallback"]: has_seg = bool(segment_ids_resolved) has_mem = bool(chunk_ids or fact_ids or tl_ids or sum_ids) if has_seg and has_mem: return "strict" if has_seg: return "partial" if has_mem: return "partial" return "fallback" async def build_chapter_bundle(self, user_id: str, chapter: Chapter) -> ChapterEvidenceBundle: notes: list[str] = [] live_segment_ids = normalize_source_segment_ids( getattr(chapter, "source_segments", None) ) row = getattr(chapter, "current_evidence_snapshot", None) row_has_closure = bool( (row and (row.segment_ids or [])) or (row and (row.memory_chunk_ids or row.memory_fact_ids or row.timeline_event_ids or row.summary_ids)) ) if ( row is not None and str(row.user_id) == str(user_id) and str(row.chapter_id) == str(chapter.id) and int(row.schema_version or 0) == EVIDENCE_SNAPSHOT_SCHEMA_VERSION and row_has_closure ): segment_ids = [ str(x) for x in (row.segment_ids or []) if str(x).strip() ] conv_ids = sorted( {str(x) for x in (row.conversation_ids or []) if str(x).strip()} ) chunk_ids = [ str(x) for x in (row.memory_chunk_ids or []) if str(x).strip() ] fact_ids = [str(x) for x in (row.memory_fact_ids or []) if str(x).strip()] tl_ids = [ str(x) for x in (row.timeline_event_ids or []) if str(x).strip() ] sum_ids = [str(x) for x in (row.summary_ids or []) if str(x).strip()] notes.extend([str(x) for x in (row.notes or []) if x]) notes.append("evidence_from_chapter_evidence_snapshot_table") tier = self._chapter_closure_tier( segment_ids_resolved=segment_ids, chunk_ids=chunk_ids, fact_ids=fact_ids, tl_ids=tl_ids, sum_ids=sum_ids, ) if live_segment_ids and set(live_segment_ids) != set(segment_ids): notes.append("live_source_segments_differ_from_snapshot_reconcile_in_pipeline") dlg = getattr(row, "message_lineage_json", None) return ChapterEvidenceBundle( user_id=user_id, chapter_id=str(chapter.id), segment_ids=segment_ids, conversation_ids=conv_ids, memory_chunk_ids=chunk_ids, memory_fact_ids=fact_ids, timeline_event_ids=tl_ids, summary_ids=sum_ids, lineage_tier=tier, notes=notes, dialogue_lineage=dlg if isinstance(dlg, dict) else None, ) snap = getattr(chapter, "evidence_bundle_json", None) snap_uid = str(snap.get("user_id") or "") if isinstance(snap, dict) else "" snap_has_closure = bool( (isinstance(snap, dict) and (snap.get("segment_ids") or [])) or ( isinstance(snap, dict) and ( snap.get("memory_chunk_ids") or snap.get("memory_fact_ids") or snap.get("timeline_event_ids") or snap.get("summary_ids") ) ) ) use_snap = ( isinstance(snap, dict) and int(snap.get("schema_version") or 0) == EVIDENCE_SNAPSHOT_SCHEMA_VERSION and str(snap.get("chapter_id") or "") == str(chapter.id) and (not snap_uid or snap_uid == str(user_id)) and snap_has_closure ) if use_snap and isinstance(snap, dict): segment_ids = [str(x) for x in (snap.get("segment_ids") or []) if str(x).strip()] conv_ids = sorted( {str(x) for x in (snap.get("conversation_ids") or []) if str(x).strip()} ) chunk_ids = [str(x) for x in (snap.get("memory_chunk_ids") or []) if str(x).strip()] fact_ids = [str(x) for x in (snap.get("memory_fact_ids") or []) if str(x).strip()] tl_ids = [str(x) for x in (snap.get("timeline_event_ids") or []) if str(x).strip()] sum_ids = [str(x) for x in (snap.get("summary_ids") or []) if str(x).strip()] notes.extend([str(x) for x in (snap.get("notes") or []) if x]) notes.append("evidence_from_chapter_evidence_bundle_json_column") tier = self._chapter_closure_tier( segment_ids_resolved=segment_ids, chunk_ids=chunk_ids, fact_ids=fact_ids, tl_ids=tl_ids, sum_ids=sum_ids, ) if live_segment_ids and set(live_segment_ids) != set(segment_ids): notes.append("live_source_segments_differ_from_snapshot_reconcile_in_pipeline") snap_dlg = snap.get("message_lineage_json") if isinstance(snap, dict) else None return ChapterEvidenceBundle( user_id=user_id, chapter_id=str(chapter.id), segment_ids=segment_ids, conversation_ids=conv_ids, memory_chunk_ids=chunk_ids, memory_fact_ids=fact_ids, timeline_event_ids=tl_ids, summary_ids=sum_ids, lineage_tier=tier, notes=notes, dialogue_lineage=snap_dlg if isinstance(snap_dlg, dict) else None, ) segment_ids = live_segment_ids if not segment_ids: notes.append("no_source_segments") notes.append("fallback_lineage_transcript_pending") return ChapterEvidenceBundle( user_id=user_id, chapter_id=str(chapter.id), segment_ids=[], conversation_ids=[], lineage_tier="fallback", notes=notes, dialogue_lineage=None, ) segments = await fetch_segments_for_user( self._db, user_id=user_id, segment_ids=segment_ids ) resolved_seg_ids = [s.id for s in segments] or segment_ids if len(segments) < len(segment_ids): notes.append("some_segments_missing_or_foreign_user") conv_ids = sorted({str(s.conversation_id) for s in segments if s.conversation_id}) chunk_ids, fact_ids, tl_ids, sum_ids = await fetch_memory_closure_for_conversations( self._db, user_id=user_id, conversation_ids=conv_ids ) tier = self._chapter_closure_tier( segment_ids_resolved=resolved_seg_ids, chunk_ids=chunk_ids, fact_ids=fact_ids, tl_ids=tl_ids, sum_ids=sum_ids, ) if tier == "partial": notes.append( "chapter_source_segments_union_semantics=partial_lineage_until_snapshot" ) elif tier == "strict": notes.append("chapter_lineage_strict_segments_plus_memory_closure") segs_ord = _segments_in_order(segments, resolved_seg_ids) dlg_live = aggregate_lineage_from_segments( segs_ord, conversation_id_fallback=conv_ids[0] if conv_ids else None ) return ChapterEvidenceBundle( user_id=user_id, chapter_id=str(chapter.id), segment_ids=resolved_seg_ids, conversation_ids=conv_ids, memory_chunk_ids=chunk_ids, memory_fact_ids=fact_ids, timeline_event_ids=tl_ids, summary_ids=sum_ids, lineage_tier=tier, notes=notes, dialogue_lineage=dlg_live, ) async def format_chapter_bundle( self, bundle: ChapterEvidenceBundle ) -> tuple[FormattedMemoirEvidence, ChapterEvidenceBundle]: """若 tier=fallback，调用方应先将要并入 transcripts 写入 session；此处只负责 segment 路径。""" if bundle.lineage_tier == "fallback": ft = await fallback_user_transcript_evidence(self._db, bundle.user_id) notes = list(bundle.notes) notes.append("used_legacy_recent_conversations_transcript") bundle = bundle.model_copy(update={"notes": notes}) formatted = format_chapter_for_judge( bundle, transcript=ft, chunks=[], facts=[], events=[], summaries=[], ) return formatted, bundle segs = await fetch_segments_for_user( self._db, user_id=bundle.user_id, segment_ids=bundle.segment_ids ) ai_map = await fetch_ai_messages_for_segments( self._db, user_id=bundle.user_id, segment_ids=[s.id for s in segs] ) transcript = build_segment_transcript(segs, ai_map) chunks = await load_chunks_by_ids( self._db, user_id=bundle.user_id, chunk_ids=bundle.memory_chunk_ids ) facts = await load_facts_by_ids( self._db, user_id=bundle.user_id, fact_ids=bundle.memory_fact_ids ) events = await load_timeline_by_ids( self._db, user_id=bundle.user_id, event_ids=bundle.timeline_event_ids ) summaries = await load_summaries_by_ids( self._db, user_id=bundle.user_id, summary_ids=bundle.summary_ids ) formatted = format_chapter_for_judge( bundle, transcript=transcript, chunks=chunks, facts=facts, events=events, summaries=summaries, ) return formatted, bundle async def build_story_bundle(self, user_id: str, story_id: str) -> StoryEvidenceBundle: st = await get_story_for_eval_trace(self._db, user_id=user_id, story_id=story_id) if not st: return StoryEvidenceBundle( user_id=user_id, story_id=story_id, lineage_tier="fallback", notes=["story_not_found"], dialogue_lineage=None, ) links = list(st.evidence_links or []) lc, lf, lt, ls = story_link_ids_by_type(links) notes: list[str] = [] chapter_ids = await list_chapter_ids_for_story( self._db, user_id=user_id, story_id=str(st.id) ) if lc or lf or lt or ls: # 结构化以 link 为准；会话级 transcript 尝试从挂靠章节 source_segments 收缩 seg_ids: list[str] = [] conv_ids: list[str] = [] for cid in chapter_ids: ch = await get_chapter_for_eval_trace( self._db, user_id=user_id, chapter_id=cid ) if not ch: continue seg_ids.extend(normalize_source_segment_ids(ch.source_segments)) # 保序去重 seen_s: set[str] = set() dedup_seg: list[str] = [] for s in seg_ids: if s not in seen_s: seen_s.add(s) dedup_seg.append(s) segments = await fetch_segments_for_user( self._db, user_id=user_id, segment_ids=dedup_seg ) conv_ids = sorted({str(s.conversation_id) for s in segments if s.conversation_id}) if dedup_seg and not segments: notes.append("chapter_segment_ids_unresolved") if conv_ids: notes.append("transcript_from_chapter_source_segments") else: notes.append("no_chapter_segments_for_transcript_context") bound_transcript = bool(segments) story_tier: Literal["strict", "partial", "fallback"] = "strict" if (lc or lf or lt or ls) and not bound_transcript: notes.append("structured_evidence_without_bound_transcript") story_tier = "partial" dlg = await self._story_dialogue_lineage(st, segments, dedup_seg) return StoryEvidenceBundle( user_id=user_id, story_id=str(st.id), segment_ids=[s.id for s in segments] or dedup_seg, conversation_ids=conv_ids, memory_chunk_ids=lc, memory_fact_ids=lf, timeline_event_ids=lt, summary_ids=ls, lineage_tier=story_tier, notes=notes, augmented_with_chapter_context=bool(chapter_ids), story_link_evidence_count=len(links), fallback_chapter_ids=chapter_ids, dialogue_lineage=dlg, ) # 无 StoryEvidenceLink：由章节 source_segments 推导 partial；再不行则 fallback seg_ids = [] conv_ids: list[str] = [] for cid in chapter_ids: ch = await get_chapter_for_eval_trace( self._db, user_id=user_id, chapter_id=cid ) if not ch: continue seg_ids.extend(normalize_source_segment_ids(ch.source_segments)) seen_s = set() dedup_seg = [] for s in seg_ids: if s not in seen_s: seen_s.add(s) dedup_seg.append(s) if dedup_seg: segments = await fetch_segments_for_user( self._db, user_id=user_id, segment_ids=dedup_seg ) conv_ids = sorted({str(s.conversation_id) for s in segments if s.conversation_id}) chunk_ids, fact_ids, tl_ids, sum_ids = ( await fetch_memory_closure_for_conversations( self._db, user_id=user_id, conversation_ids=conv_ids ) ) notes.append("fallback_lineage_no_story_evidence_links") notes.append("augmented_with_chapter_context") dlg2 = await self._story_dialogue_lineage(st, segments, dedup_seg) return StoryEvidenceBundle( user_id=user_id, story_id=str(st.id), segment_ids=[s.id for s in segments] or dedup_seg, conversation_ids=conv_ids, memory_chunk_ids=chunk_ids, memory_fact_ids=fact_ids, timeline_event_ids=tl_ids, summary_ids=sum_ids, lineage_tier="partial", notes=notes, augmented_with_chapter_context=True, story_link_evidence_count=0, fallback_chapter_ids=chapter_ids, dialogue_lineage=dlg2, ) notes.append("no_story_evidence_links_and_no_chapter_segments") notes.append("fallback_lineage_transcript_pending") dlg3 = await self._story_dialogue_lineage(st, [], []) return StoryEvidenceBundle( user_id=user_id, story_id=str(st.id), lineage_tier="fallback", notes=notes, story_link_evidence_count=0, fallback_chapter_ids=chapter_ids, dialogue_lineage=dlg3, ) async def format_story_bundle( self, bundle: StoryEvidenceBundle ) -> tuple[FormattedMemoirEvidence, StoryEvidenceBundle]: if bundle.lineage_tier == "fallback": ft = await fallback_user_transcript_evidence(self._db, bundle.user_id) notes = list(bundle.notes) notes.append("used_legacy_recent_conversations_transcript") bundle = bundle.model_copy(update={"notes": notes}) formatted = format_story_for_judge( bundle, transcript=ft, chunks=[], facts=[], events=[], summaries=[], ) return formatted, bundle segs = await fetch_segments_for_user( self._db, user_id=bundle.user_id, segment_ids=bundle.segment_ids ) ai_map = await fetch_ai_messages_for_segments( self._db, user_id=bundle.user_id, segment_ids=[s.id for s in segs] ) transcript = build_segment_transcript(segs, ai_map) chunks = await load_chunks_by_ids( self._db, user_id=bundle.user_id, chunk_ids=bundle.memory_chunk_ids ) facts = await load_facts_by_ids( self._db, user_id=bundle.user_id, fact_ids=bundle.memory_fact_ids ) events = await load_timeline_by_ids( self._db, user_id=bundle.user_id, event_ids=bundle.timeline_event_ids ) summaries = await load_summaries_by_ids( self._db, user_id=bundle.user_id, summary_ids=bundle.summary_ids ) formatted = format_story_for_judge( bundle, transcript=transcript, chunks=chunks, facts=facts, events=events, summaries=summaries, ) return formatted, bundle