fix(rag): retrieve complete course assignments

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

@@ -365,8 +365,19 @@ async def _feishu_download_async(path: str, config: FeishuRetrievalConfig) -> by
 
 
 async def _get_docx_content_async(doc_token: str, config: FeishuRetrievalConfig) -> str:
-    data = await _feishu_get_async(f"/open-apis/docx/v1/documents/{doc_token}/blocks", config=config)
-    items = data.get("data", {}).get("items", [])
+    path = f"/open-apis/docx/v1/documents/{doc_token}/blocks"
+    items: list[dict] = []
+    page_token = ""
+    while True:
+        params: dict[str, Any] = {"page_size": 500}
+        if page_token:
+            params["page_token"] = page_token
+        data = await _feishu_get_async(path, params=params, config=config)
+        page = data.get("data", {})
+        items.extend(page.get("items", []))
+        page_token = str(page.get("page_token") or "")
+        if not page.get("has_more") or not page_token:
+            break
     return _parse_docx_content(items)
 
 

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

@@ -234,8 +234,19 @@ def _parse_docx_content(blocks: list[dict]) -> str:
 
 def _get_docx_content(doc_token: str, config: FeishuRetrievalConfig) -> str:
     """读取单个飞书云文档内容"""
-    data = _feishu_get(f"/open-apis/docx/v1/documents/{doc_token}/blocks", config=config)
-    items = data.get("data", {}).get("items", [])
+    path = f"/open-apis/docx/v1/documents/{doc_token}/blocks"
+    items: list[dict] = []
+    page_token = ""
+    while True:
+        params: dict[str, Any] = {"page_size": 500}
+        if page_token:
+            params["page_token"] = page_token
+        data = _feishu_get(path, params=params, config=config)
+        page = data.get("data", {})
+        items.extend(page.get("items", []))
+        page_token = str(page.get("page_token") or "")
+        if not page.get("has_more") or not page_token:
+            break
     return _parse_docx_content(items)
 
 

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

@@ -32,6 +32,7 @@ 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
 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"}
 
 
@@ -107,13 +108,22 @@ class KnowledgeAgentService:
                 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)
-                selected = cls._select_sections(candidates)
+                practice_overview = cls._is_practice_overview(retrieval_question)
+                selected = cls._select_sections(
+                    candidates,
+                    limit=cls._selection_limit(retrieval_question),
+                    prefer_numbered_practice=practice_overview,
+                )
+                selected_contents = {
+                    item.section.id: cls._protect_content(cls._read_complete_section(db, item.section))
+                    for item in selected
+                }
                 chunks = [
                     RetrievedChunk(
                         knowledge_id=item.knowledge.id,
                         knowledge_name=item.knowledge.name,
                         title=item.section.title,
-                        content=cls._protect_content(item.section.content),
+                        content=selected_contents[item.section.id],
                         version_id=item.version.id,
                         section_id=item.section.id,
                         chunk_id=item.chunk.id,
@@ -121,7 +131,7 @@ class KnowledgeAgentService:
                     )
                     for item in selected
                 ]
-                trace.append(cls._trace("read_knowledge_sections", len(trace) + 1, {"sectionIds": [item.section.id for item in selected]}, {"count": len(selected), "sections": [{"knowledgeId": x.knowledge.id, "knowledgeName": x.knowledge.name, "sectionId": x.section.id, "title": x.section.title, "content": cls._protect_content(x.section.content), "adopted": True} for x in selected]}, started))
+                trace.append(cls._trace("read_knowledge_sections", len(trace) + 1, {"sectionIds": [item.section.id for item in selected]}, {"count": len(selected), "sections": [{"knowledgeId": x.knowledge.id, "knowledgeName": x.knowledge.name, "sectionId": x.section.id, "title": x.section.title, "content": selected_contents[x.section.id], "adopted": True} for x in selected]}, started))
                 cls._persist_candidates(db, log.id, candidates)
                 log.knowledge_called = 1
                 log.selected_knowledge_ids = ",".join(str(item) for item in selected_ids)
@@ -328,13 +338,14 @@ class KnowledgeAgentService:
             section, knowledge, version = sections.get(chunk.section_id), knowledge_rows.get(chunk.knowledge_id), versions.get(chunk.version_id)
             if not section or not knowledge or not version:
                 continue
-            haystack = f"{chunk.title} {chunk.normalized_text} {chunk.content}".lower()
+            haystack = f"{knowledge.name} {chunk.title} {chunk.normalized_text} {chunk.content}".lower()
             score = 0.0
             for term in terms:
                 count = haystack.count(term)
                 if count:
                     idf = math.log((len(chunks) + 1) / (frequencies.get(term, 0) + 1)) + 1
                     score += (count / (count + 1.2)) * idf * (2.4 if term in chunk.title.lower() else 1.0)
+            score += cls._title_intent_boost(chunk.title, terms)
             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)
@@ -367,7 +378,21 @@ class KnowledgeAgentService:
             trace.append(cls._trace("Rerank", len(trace) + 1, {"candidateCount": len(candidates)}, {"count": len(candidates), "mode": "lexical_fallback"}, started, status="fallback", error=str(exc)[:300]))
 
     @staticmethod
-    def _select_sections(candidates: list[Candidate]) -> list[Candidate]:
+    def _select_sections(
+        candidates: list[Candidate],
+        *,
+        limit: int = MAX_SELECTED_SECTIONS,
+        prefer_numbered_practice: bool = False,
+    ) -> list[Candidate]:
+        if prefer_numbered_practice:
+            candidates = sorted(
+                candidates,
+                key=lambda item: (
+                    KnowledgeAgentService._is_numbered_practice_title(item.chunk.title),
+                    item.rerank_score if item.rerank_score is not None else item.lexical_score,
+                ),
+                reverse=True,
+            )
         selected: list[Candidate] = []
         seen: set[int] = set()
         for item in candidates:
@@ -376,7 +401,7 @@ class KnowledgeAgentService:
                 item.discard_reason = "相关性不足"
             elif item.section.id in seen:
                 item.discard_reason = "同一父章节已有更高分候选"
-            elif len(selected) >= MAX_SELECTED_SECTIONS:
+            elif len(selected) >= limit:
                 item.discard_reason = "超过本轮章节数量限制"
             else:
                 item.selected = True
@@ -384,6 +409,54 @@ class KnowledgeAgentService:
                 seen.add(item.section.id)
         return selected
 
+    @staticmethod
+    def _selection_limit(question: str) -> int:
+        if KnowledgeAgentService._is_practice_overview(question):
+            return MAX_OVERVIEW_SELECTED_SECTIONS
+        return MAX_SELECTED_SECTIONS
+
+    @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")
+        practice_markers = ("作业", "功课", "练习")
+        return any(marker in question for marker in practice_markers) and any(marker in question for marker in overview_markers)
+
+    @staticmethod
+    def _is_numbered_practice_title(title: str) -> bool:
+        return bool(re.search(r"(?:练习|作业|功课)\s*[一二三四五六七八九十百\d]+", title))
+
+    @staticmethod
+    def _title_intent_boost(title: str, terms: list[str]) -> float:
+        if not {"作业", "功课", "练习"}.intersection(terms):
+            return 0.0
+        if KnowledgeAgentService._is_numbered_practice_title(title):
+            return 20.0
+        if any(marker in title for marker in ("练习", "作业", "功课")):
+            return 3.0
+        return 0.0
+
+    @staticmethod
+    def _read_complete_section(db: Session, section: KnowledgeSection) -> str:
+        """Read a heading together with its lower-level child sections."""
+        level = _heading_level(section.content)
+        if level is None:
+            return section.content
+        rows = db.scalars(
+            select(KnowledgeSection)
+            .where(
+                KnowledgeSection.version_id == section.version_id,
+                KnowledgeSection.sort_order >= section.sort_order,
+            )
+            .order_by(KnowledgeSection.sort_order)
+        ).all()
+        parts: list[str] = []
+        for row in rows:
+            row_level = _heading_level(row.content)
+            if row.id != section.id and row_level is not None and row_level <= level:
+                break
+            parts.append(row.content)
+        return "\n\n".join(parts)
+
     @staticmethod
     def _persist_candidates(db: Session, log_id: int, candidates: list[Candidate]) -> None:
         for item in candidates:
@@ -443,6 +516,11 @@ def _json_trace_value(value):
     return str(value)
 
 
+def _heading_level(content: str) -> int | None:
+    match = re.match(r"^(#{1,6})\s+", content.lstrip())
+    return len(match.group(1)) if match else None
+
+
 def _extract_json(value: str) -> str:
     match = re.search(r"\{[\s\S]*\}", value)
     if not match:

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

@@ -455,6 +455,9 @@ _SYNONYMS = {
     "焦虑": ["担心", "紧张", "害怕"],
     "沟通": ["交流", "对话", "聊天"],
     "课程": ["训练营", "大本营", "课堂"],
+    "作业": ["功课", "练习"],
+    "功课": ["作业", "练习"],
+    "练习": ["作业", "功课"],
 }
 
 

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

@@ -121,7 +121,7 @@ class PromptService:
 
         # 知识库上下文
         context = "\n\n".join(
-            f"[已回读完整章节 {index}] {chunk.title}\n{chunk.content}"
+            f"[已回读完整章节 {index}] {chunk.title}\n来源知识库：{chunk.knowledge_name}\n{chunk.content}"
             for index, chunk in enumerate(chunks, start=1)
         )
         if not context:
@@ -152,7 +152,21 @@ class PromptService:
             if content and message.role in {"user", "assistant"}:
                 messages.append({"role": message.role, "content": content})
 
-        messages.append({"role": "system", "content": f"[本轮可靠知识上下文]\n{context}"})
+        messages.append({
+            "role": "system",
+            "content": (
+                "[本轮可靠知识上下文]\n"
+                "以下章节均来自当前正式开放、已发布的知识库，可以作为本轮回答的可靠依据。"
+                "如果其中已经明确包含用户询问的课程、作业、功课或练习，不得再以‘没有资料’、"
+                "‘无法确认’或要求用户补充课程名称来回避回答；应直接根据章节标题和正文归纳。"
+                "用户问‘某课程的作业是什么’时，含义是询问该课程包含哪些作业或练习，"
+                "不要求知识库中必须存在一个与‘某课程作业’完全同名的章节。"
+                "回答作业清单类问题时，应直接给出作业名称，并概括每项作业的用途或核心方向；"
+                "正文已经提供基本步骤时可以简要说明，不要把本轮已经取到的内容留到下一轮再问。"
+                "章节标题可以作为作业或练习名称，标题下的子章节是对应的正式说明。\n\n"
+                f"{context}"
+            ),
+        })
         messages.append({"role": "user", "content": question.strip()})
         return messages