chore: sync local code

backend/.env.example

@@ -27,6 +27,8 @@ LANGSMITH_TRACING=false
 # 默认建议创建独立数据库，如 ctb；若直接使用 postgres 默认库，也可改为 /postgres。
 # APP_DATABASE_URL=postgresql+psycopg://postgres:root@localhost:15432/ctb
 # APP_DATABASE_ECHO=false
+# APP_POSTGRES_VECTOR_DIMENSIONS 必须和所选 embedding 模型维度一致。
+# 例如 text-embedding-3-small 默认是 1536 维。
 # APP_POSTGRES_VECTOR_DIMENSIONS=1536
 
 # ============================================================================

backend/core/config.py

@@ -205,6 +205,7 @@ class Settings(BaseSettings):
     db_path: Path | None = None
     database_url: str = ""
     database_echo: bool = False
+    postgres_vector_dimensions: int = 1536
 
     # 各类子目录（由 validator 从 runtime_dir 派生，可独立覆盖以便测试）
     upload_dir: Path | None = None
@@ -225,7 +226,7 @@ class Settings(BaseSettings):
         True  # 是否信任系统代理环境变量，Windows 下设为 False 可解 WinError 10054
     )
 
-    rag_embedding_model: str = "text-embedding-v3"
+    rag_embedding_model: str = "text-embedding-3-small"
     rag_embedding_api_key: str = ""
     rag_embedding_base_url: str = ""
 

backend/core/model_selection.py

@@ -22,9 +22,11 @@ class LLMSelectionError(Exception):
 
 
 def split_models(model_name: str | None) -> list[str]:
-    """返回模型名称列表。现在只支持单个模型，但保留列表形式以兼容调用方。"""
-    name = (model_name or "").strip()
-    return [name] if name else []
+    """返回模型名称列表，兼容逗号分隔的历史存储格式。"""
+    raw = str(model_name or "").strip()
+    if not raw:
+        return []
+    return [item.strip() for item in raw.split(",") if item.strip()]
 
 
 def build_managed_provider_context(db):

backend/core/rag.py

@@ -317,56 +317,48 @@ def index_question(db: Session, question_id: int) -> bool:
 
     vector_json = _serialize_vector(vector)
 
-    if existing:
-        existing.content = chunk_data["content"]
-        existing.metadata_json = json.dumps(chunk_data["metadata"], ensure_ascii=False)
-        existing.content_hash = chunk_data["content_hash"]
-        existing.embedding_model = settings.rag_embedding_model if vector else None
-        existing.vector_json = vector_json
-    else:
-        chunk = RagDocumentChunk(
-            user_id=question.user_id,
-            project_id=question.project_id,
-            source_type="question",
-            source_id=question_id,
-            chunk_index=0,
-            content=chunk_data["content"],
-            metadata_json=json.dumps(chunk_data["metadata"], ensure_ascii=False),
-            content_hash=chunk_data["content_hash"],
-            embedding_model=settings.rag_embedding_model if vector else None,
-            vector_json=vector_json,
-        )
-        db.add(chunk)
-
     try:
-        db.flush()
-        target_chunk = existing or chunk
-        _write_postgres_vector(db, target_chunk.id, vector)
-        db.commit()
+        with db.begin_nested():
+            if existing:
+                existing.content = chunk_data["content"]
+                existing.metadata_json = json.dumps(chunk_data["metadata"], ensure_ascii=False)
+                existing.content_hash = chunk_data["content_hash"]
+                existing.embedding_model = settings.rag_embedding_model if vector else None
+                existing.vector_json = vector_json
+                target_chunk = existing
+            else:
+                target_chunk = RagDocumentChunk(
+                    user_id=question.user_id,
+                    project_id=question.project_id,
+                    source_type="question",
+                    source_id=question_id,
+                    chunk_index=0,
+                    content=chunk_data["content"],
+                    metadata_json=json.dumps(chunk_data["metadata"], ensure_ascii=False),
+                    content_hash=chunk_data["content_hash"],
+                    embedding_model=settings.rag_embedding_model if vector else None,
+                    vector_json=vector_json,
+                )
+                db.add(target_chunk)
+
+            db.flush()
+            _write_postgres_vector(db, target_chunk.id, vector)
         return True
     except Exception as e:
-        db.rollback()
         logger.error("索引题目 %d 失败: %s", question_id, e)
         return False
 
 
 def delete_question_chunks(db: Session, question_id: int) -> int:
     """删除错题关联的所有 RAG chunk"""
-    count = (
+    return (
         db.query(RagDocumentChunk)
         .filter(
             RagDocumentChunk.source_type == "question",
             RagDocumentChunk.source_id == question_id,
         )
         .delete(synchronize_session=False)
     )
-    try:
-        db.commit()
-    except Exception as e:
-        db.rollback()
-        logger.error("删除题目 %d 的 chunk 失败: %s", question_id, e)
-        return 0
-    return count
 
 
 # ---------------------------------------------------------------------------

backend/db/__init__.py

@@ -2,7 +2,7 @@
 数据库模块：引擎创建、Session 工厂、初始化函数
 """
 
-from sqlalchemy import create_engine, event
+from sqlalchemy import create_engine, event, inspect, text
 from sqlalchemy.orm import sessionmaker
 import os
 import sys
@@ -154,7 +154,46 @@ def _migrate_schema():
         conn.close()
 
 
+def _prepare_postgresql_extensions():
+    """在 PostgreSQL 建表前确保 pgvector 扩展可用。"""
+    if not is_postgresql_backend(engine):
+        return
+    with engine.begin() as conn:
+        conn.execute(text("CREATE EXTENSION IF NOT EXISTS vector"))
+
+
+def _ensure_postgresql_schema():
+    """为 PostgreSQL 现有表补齐 pgvector 列和索引。"""
+    if not is_postgresql_backend(engine):
+        return
+
+    inspector = inspect(engine)
+    table_names = set(inspector.get_table_names())
+    if "rag_document_chunks" not in table_names:
+        return
+
+    columns = {column["name"] for column in inspector.get_columns("rag_document_chunks")}
+    with engine.begin() as conn:
+        conn.execute(text("CREATE EXTENSION IF NOT EXISTS vector"))
+        if "embedding_vector" not in columns:
+            conn.execute(
+                text(
+                    "ALTER TABLE rag_document_chunks "
+                    f"ADD COLUMN embedding_vector vector({settings.postgres_vector_dimensions})"
+                )
+            )
+        conn.execute(
+            text(
+                "CREATE INDEX IF NOT EXISTS idx_rag_document_chunks_embedding_vector "
+                "ON rag_document_chunks USING ivfflat (embedding_vector vector_cosine_ops) "
+                "WITH (lists = 100)"
+            )
+        )
+
+
 def init_db():
     """初始化数据库：建表并执行轻量级自动迁移"""
+    _prepare_postgresql_extensions()
     Base.metadata.create_all(bind=engine)
     _migrate_schema()
+    _ensure_postgresql_schema()

backend/db/crud/projects.py

@@ -6,7 +6,7 @@
 
 from db.models import (
     ChatSession, Note, NoteTagMapping, Project, Question,
-    QuestionEmbedding, QuestionTagMapping, UploadBatch,
+    QuestionEmbedding, QuestionTagMapping, RagDocumentChunk, UploadBatch,
 )
 
 
@@ -141,6 +141,10 @@ def delete_project(db: Session, project_id: int, user_id=None) -> bool:
     # 先删除题目和笔记的关联子表，再删除题目/笔记本身
     question_ids = [q.id for q in db.query(Question.id).filter(Question.project_id == project.id).all()]
     if question_ids:
+        db.query(RagDocumentChunk).filter(
+            RagDocumentChunk.source_type == "question",
+            RagDocumentChunk.source_id.in_(question_ids),
+        ).delete(synchronize_session=False)
         db.query(QuestionEmbedding).filter(QuestionEmbedding.question_id.in_(question_ids)).delete(synchronize_session=False)
         db.query(ChatSession).filter(ChatSession.question_id.in_(question_ids)).delete(synchronize_session=False)
         db.query(QuestionTagMapping).filter(QuestionTagMapping.question_id.in_(question_ids)).delete(synchronize_session=False)

backend/db/crud/questions.py

@@ -358,8 +358,6 @@ def save_questions_to_db(
     # 科目由编排智能体识别，不再使用关键词匹配
     subject = batch_info.get("subject") or "未知"
     project_id = project_id or batch_info.get("project_id")
-    if not project_id:
-        raise ValueError("PROJECT_REQUIRED")
 
     # 创建批次记录
     batch = UploadBatch(
@@ -693,6 +691,10 @@ def delete_question(db: Session, question_id: int, user_id=None) -> bool:
 
     try:
         batch_id = question.batch_id
+        from core.rag import delete_question_chunks
+
+        delete_question_chunks(db, question_id)
+
         # 删除关联的标签映射
         db.query(QuestionTagMapping).filter(QuestionTagMapping.question_id == question_id).delete()
 

backend/db/models.py

@@ -4,13 +4,32 @@
 
 import uuid
 from sqlalchemy import Column, Integer, String, Text, DateTime, Boolean, ForeignKey, UniqueConstraint
-from sqlalchemy.orm import relationship
-from sqlalchemy.ext.declarative import declarative_base
+from sqlalchemy.types import TypeDecorator
+from sqlalchemy.orm import declarative_base, relationship
 from datetime import datetime
 
+from core.config import settings
+
+try:
+    from pgvector.sqlalchemy import Vector
+except ImportError:  # pragma: no cover - optional on SQLite-only setups
+    Vector = None
+
 Base = declarative_base()
 
 
+class EmbeddingVectorType(TypeDecorator):
+    """Use pgvector on PostgreSQL and plain text elsewhere."""
+
+    impl = Text
+    cache_ok = True
+
+    def load_dialect_impl(self, dialect):
+        if dialect.name == "postgresql" and Vector is not None:
+            return dialect.type_descriptor(Vector(settings.postgres_vector_dimensions))
+        return dialect.type_descriptor(Text())
+
+
 class User(Base):
     """用户表"""
     __tablename__ = "users"
@@ -330,5 +349,6 @@ class RagDocumentChunk(Base):
     content_hash = Column(String(64), default="", index=True)
     embedding_model = Column(String(100), nullable=True)
     vector_json = Column(Text, nullable=True)
+    embedding_vector = Column(EmbeddingVectorType(), nullable=True)
     created_at = Column(DateTime, default=datetime.utcnow, index=True)
     updated_at = Column(DateTime, default=datetime.utcnow, onupdate=datetime.utcnow)

backend/routes/chat.py

@@ -412,21 +412,6 @@ def stream_chat(session_id):
 
         user_id = session.get("user_id")
         with SessionLocal() as db:
-            try:
-                selection = resolve_llm_selection(
-                    db,
-                    user_id=user_id,
-                    category=model_provider,
-                    model_name=model_name,
-                    provider_source=provider_source,
-                    provider_id=provider_id,
-                )
-            except LLMSelectionError as e:
-                return (
-                    jsonify({"success": False, "code": e.code, "error": e.message}),
-                    e.status_code,
-                )
-
             uid = _effective_user_id()
             cs_query = (
                 db.query(ChatSessionModel)
@@ -446,6 +431,21 @@ def stream_chat(session_id):
             if not chat_session:
                 return jsonify({"success": False, "error": "对话不存在"}), 404
 
+            try:
+                selection = resolve_llm_selection(
+                    db,
+                    user_id=user_id,
+                    category=model_provider,
+                    model_name=model_name,
+                    provider_source=provider_source,
+                    provider_id=provider_id,
+                )
+            except LLMSelectionError as e:
+                return (
+                    jsonify({"success": False, "code": e.code, "error": e.message}),
+                    e.status_code,
+                )
+
             should_consume_quota = bool(user_id) and uses_server_llm_selection(
                 selection["source"],
                 db=db,

backend/routes/questions.py

@@ -456,8 +456,10 @@ def rag_reindex():
                 with SessionLocal() as db:
                     success = index_question(db, qid)
                     if success:
+                        db.commit()
                         indexed += 1
                     else:
+                        db.rollback()
                         skipped += 1
             except Exception as e:
                 logger.warning("索引题目 %d 失败: %s", qid, e)
@@ -531,7 +533,10 @@ def update_question(question_id):
         try:
             from core.rag import index_question
             with SessionLocal() as db:
-                index_question(db, question_id)
+                if index_question(db, question_id):
+                    db.commit()
+                else:
+                    db.rollback()
         except Exception as e:
             logger.warning(f"更新题目 {question_id} 的 RAG 索引失败: {e}")
 
@@ -560,7 +565,10 @@ def update_question_answer(question_id):
         try:
             from core.rag import index_question
             with SessionLocal() as db:
-                index_question(db, question_id)
+                if index_question(db, question_id):
+                    db.commit()
+                else:
+                    db.rollback()
         except Exception as e:
             logger.warning(f"更新题目 {question_id} 的用户作答索引失败: {e}")
 
@@ -624,6 +632,8 @@ def save_to_db():
 
         if not isinstance(selected_uids, list) or not selected_uids:
             return jsonify({'success': False, 'error': '请选择至少一道题目'}), 400
+        if project_id is None:
+            return jsonify({'success': False, 'error': '请先创建并选择一个错题库'}), 400
 
         run_id = data.get('run_id')
         record_id = data.get('record_id')
@@ -677,15 +687,11 @@ def save_to_db():
 
         with SessionLocal() as db:
             try:
-                project_id = (
-                    crud.require_project_id(
-                        db,
-                        project_id,
-                        user_id=session.get('user_id'),
-                        project_type="question",
-                    )
-                    if project_id
-                    else None
+                project_id = crud.resolve_project_id(
+                    db,
+                    project_id,
+                    user_id=session.get('user_id'),
+                    project_type="question",
                 )
             except ValueError as exc:
                 if str(exc) == "PROJECT_REQUIRED":
@@ -878,7 +884,10 @@ def save_question_answer(question_id):
         try:
             from core.rag import index_question
             with SessionLocal() as db:
-                index_question(db, question_id)
+                if index_question(db, question_id):
+                    db.commit()
+                else:
+                    db.rollback()
         except Exception as e:
             logger.warning(f"更新题目 {question_id} 的 RAG 索引失败: {e}")