fix(rag): recall complete course homework lists

ai_knowledge_base_v2/apps/backend/app/services/knowledge_agent_service.py

@@ -32,7 +32,8 @@ from app.services.rag_service import PromptService, RagResult, RetrievedChunk
 SAFETY_RULE_VERSION = "minimum-safety-v1"
 MAX_LEXICAL_CANDIDATES = 12
 MAX_SELECTED_SECTIONS = 4
-MAX_OVERVIEW_SELECTED_SECTIONS = 8
+MAX_OVERVIEW_LEXICAL_CANDIDATES = 48
+MAX_OVERVIEW_SELECTED_SECTIONS = 24
 BUSINESS_MARKERS = {"课程", "大本营", "训练营", "老师", "卢慧", "功课", "学员", "课堂", "练习", "觉察", "内在"}
 
 
@@ -105,10 +106,17 @@ class KnowledgeAgentService:
         try:
             if need_knowledge and selected_ids:
                 terms = cls._query_terms(retrieval_question)
-                candidates = cls.search_knowledge(db, terms, selected_ids, catalog)
-                trace.append(cls._trace("search_knowledge", len(trace) + 1, {"queryTerms": terms, "knowledgeIds": selected_ids}, {"candidateCount": len(candidates), "candidates": [cls._candidate_trace(x) for x in candidates]}, started))
-                await cls._rerank(db, retrieval_question, candidates, trace, started)
                 practice_overview = cls._is_practice_overview(retrieval_question)
+                candidate_limit = cls._candidate_limit(retrieval_question)
+                candidates = cls.search_knowledge(
+                    db,
+                    terms,
+                    selected_ids,
+                    catalog,
+                    candidate_limit=candidate_limit,
+                )
+                trace.append(cls._trace("search_knowledge", len(trace) + 1, {"queryTerms": terms, "knowledgeIds": selected_ids, "candidateLimit": candidate_limit}, {"candidateCount": len(candidates), "candidates": [cls._candidate_trace(x) for x in candidates]}, started))
+                await cls._rerank(db, retrieval_question, candidates, trace, started)
                 selected = cls._select_sections(
                     candidates,
                     limit=cls._selection_limit(retrieval_question),
@@ -323,7 +331,15 @@ class KnowledgeAgentService:
         return {"knowledgeId": item.knowledge.id, "knowledgeName": item.knowledge.name, "versionId": item.version.id, "chunkId": item.chunk.id, "sectionId": item.section.id, "title": item.chunk.title, "lexicalScore": item.lexical_score, "rerankScore": item.rerank_score, "selected": item.selected, "discardReason": item.discard_reason}
 
     @classmethod
-    def search_knowledge(cls, db: Session, terms: list[str], selected_ids: list[int], catalog: list[dict]) -> list[Candidate]:
+    def search_knowledge(
+        cls,
+        db: Session,
+        terms: list[str],
+        selected_ids: list[int],
+        catalog: list[dict],
+        *,
+        candidate_limit: int = MAX_LEXICAL_CANDIDATES,
+    ) -> list[Candidate]:
         versions_by_kb = {item["knowledgeId"]: item["versionId"] for item in catalog}
         version_ids = [versions_by_kb[item] for item in selected_ids if item in versions_by_kb]
         if not version_ids:
@@ -349,7 +365,7 @@ class KnowledgeAgentService:
             if score > 0:
                 candidates.append(Candidate(chunk, section, knowledge, version, score))
         candidates.sort(key=lambda item: (item.lexical_score, item.version.published_at or item.version.created_at), reverse=True)
-        return candidates[:MAX_LEXICAL_CANDIDATES]
+        return candidates[:candidate_limit]
 
     @classmethod
     async def _rerank(cls, db: Session, question: str, candidates: list[Candidate], trace: list[dict], started: float) -> None:
@@ -385,6 +401,15 @@ class KnowledgeAgentService:
         prefer_numbered_practice: bool = False,
     ) -> list[Candidate]:
         if prefer_numbered_practice:
+            numbered_candidates = [
+                item for item in candidates
+                if KnowledgeAgentService._is_numbered_practice_title(item.chunk.title)
+            ]
+            if numbered_candidates:
+                for item in candidates:
+                    if not KnowledgeAgentService._is_numbered_practice_title(item.chunk.title):
+                        item.discard_reason = "作业清单优先采用编号练习章节"
+                candidates = numbered_candidates
             candidates = sorted(
                 candidates,
                 key=lambda item: (
@@ -415,6 +440,12 @@ class KnowledgeAgentService:
             return MAX_OVERVIEW_SELECTED_SECTIONS
         return MAX_SELECTED_SECTIONS
 
+    @staticmethod
+    def _candidate_limit(question: str) -> int:
+        if KnowledgeAgentService._is_practice_overview(question):
+            return MAX_OVERVIEW_LEXICAL_CANDIDATES
+        return MAX_LEXICAL_CANDIDATES
+
     @staticmethod
     def _is_practice_overview(question: str) -> bool:
         overview_markers = ("有哪些", "是什么", "包括什么", "都有什么", "列出", "汇总", "总结")

ai_knowledge_base_v2/apps/backend/tests/test_knowledge_agent.py

@@ -3,6 +3,7 @@ from __future__ import annotations
 import asyncio
 import json
 from datetime import datetime
+from types import SimpleNamespace
 
 from sqlalchemy import create_engine, select
 from sqlalchemy.orm import Session
@@ -19,7 +20,7 @@ from app.models.knowledge import (
     KnowledgeVersion,
 )
 from app.models.chat import ChatMessage
-from app.services.knowledge_agent_service import KnowledgeAgentService
+from app.services.knowledge_agent_service import Candidate, KnowledgeAgentService
 
 
 def _database() -> Session:
@@ -187,13 +188,45 @@ def test_homework_overview_expands_practice_terms_and_section_limit():
 
     assert "功课" in terms
     assert "练习" in terms
-    assert KnowledgeAgentService._selection_limit("合一的作业是什么？") == 8
+    assert KnowledgeAgentService._candidate_limit("合一的作业是什么？") == 48
+    assert KnowledgeAgentService._selection_limit("合一的作业是什么？") == 24
+    assert KnowledgeAgentService._candidate_limit("这个练习怎么做？") == 12
     assert KnowledgeAgentService._selection_limit("这个练习怎么做？") == 4
     assert KnowledgeAgentService._title_intent_boost("十七、练习一：风铃式静心", terms) == 20.0
     assert KnowledgeAgentService._title_intent_boost("完整练习的方向", terms) == 3.0
     assert KnowledgeAgentService._title_intent_boost("课程定位", terms) == 0.0
 
 
+def test_homework_overview_keeps_all_numbered_practices_before_selection():
+    candidates = [
+        Candidate(
+            chunk=SimpleNamespace(title=f"练习{index}：课程作业", id=index),
+            section=SimpleNamespace(id=index),
+            knowledge=SimpleNamespace(id=1),
+            version=SimpleNamespace(id=1),
+            lexical_score=float(30 - index),
+        )
+        for index in range(1, 15)
+    ]
+
+    unrelated = Candidate(
+        chunk=SimpleNamespace(title="课程介绍", id=99),
+        section=SimpleNamespace(id=99),
+        knowledge=SimpleNamespace(id=1),
+        version=SimpleNamespace(id=1),
+        lexical_score=100.0,
+    )
+
+    selected = KnowledgeAgentService._select_sections(
+        [unrelated, *candidates],
+        limit=KnowledgeAgentService._selection_limit("原生里的作业内容都有什么"),
+        prefer_numbered_practice=True,
+    )
+
+    assert len(selected) == 14
+    assert all(item.chunk.title.startswith("练习") for item in selected)
+
+
 def test_complete_section_includes_child_headings_but_stops_at_next_peer():
     with _database() as db:
         knowledge = _add_published_knowledge(db, knowledge_id=1, name="合一课程")