feat(embedding): support multi-turn conversation context for table RAG

- Add MULTI_TURN_EMBEDDING_ENABLED and MULTI_TURN_HISTORY_COUNT config
- Add get_chat_history_questions() to retrieve recent questions from same chat
- Add build_context_query() to concatenate history questions with current question
- Update calc_table_embedding() to use context query for better table matching
- Pass history_questions through get_table_schema() to LLMService

This improves table structure retrieval accuracy by considering the full
conversation context instead of just the latest question.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <[email protected]>

Author: WainWong

backend/apps/chat/curd/chat.py

@@ -859,3 +859,36 @@ def get_old_questions(session: SessionDep, datasource: int):
     for r in result:
         records.append(r.question)
     return records
+
+
+def get_chat_history_questions(session: SessionDep, chat_id: int, limit: int = 3) -> List[str]:
+    """
+    获取当前chat的历史问题列表（按时间正序，最旧的在前）
+
+    Args:
+        session: 数据库会话
+        chat_id: 当前对话ID
+        limit: 获取的历史问题数量
+
+    Returns:
+        历史问题列表，按时间正序排列
+    """
+    stmt = (
+        select(ChatRecord.question)
+        .where(
+            and_(
+                ChatRecord.chat_id == chat_id,
+                ChatRecord.question.isnot(None),
+                ChatRecord.question != '',
+                ChatRecord.error.is_(None)
+            )
+        )
+        .order_by(ChatRecord.create_time.desc())
+        .limit(limit)
+    )
+
+    result = session.execute(stmt)
+    questions = [row.question for row in result if row.question and row.question.strip()]
+
+    # 反转列表，使最旧的在前
+    return list(reversed(questions))

backend/apps/chat/task/llm.py

@@ -31,7 +31,7 @@
     get_old_questions, save_analysis_predict_record, rename_chat, get_chart_config, \
     get_chat_chart_data, list_generate_sql_logs, list_generate_chart_logs, start_log, end_log, \
     get_last_execute_sql_error, format_json_data, format_chart_fields, get_chat_brief_generate, get_chat_predict_data, \
-    get_chat_chart_config
+    get_chat_chart_config, get_chat_history_questions
 from apps.chat.models.chat_model import ChatQuestion, ChatRecord, Chat, RenameChat, ChatLog, OperationEnum, \
     ChatFinishStep, AxisObj
 from apps.data_training.curd.data_training import get_training_template
@@ -101,6 +101,12 @@ def __init__(self, session: Session, current_user: CurrentUser, chat_question: C
         chat: Chat | None = session.get(Chat, chat_id)
         if not chat:
             raise SingleMessageError(f"Chat with id {chat_id} not found")
+
+        # 获取历史问题（用于多轮对话embedding）
+        history_questions = []
+        if settings.MULTI_TURN_EMBEDDING_ENABLED:
+            history_questions = get_chat_history_questions(session, chat_id, settings.MULTI_TURN_HISTORY_COUNT)
+
         ds: CoreDatasource | AssistantOutDsSchema | None = None
         if chat.datasource:
             # Get available datasource

backend/apps/datasource/embedding/table_embedding.py

@@ -3,14 +3,46 @@
 import json
 import time
 import traceback
+from typing import List
 
 from apps.ai_model.embedding import EmbeddingModelCache
 from apps.datasource.embedding.utils import cosine_similarity
 from common.core.config import settings
 from common.utils.utils import SQLBotLogUtil
 
 
-def get_table_embedding(tables: list[dict], question: str):
+def build_context_query(current_question: str, history_questions: List[str] = None) -> str:
+    """
+    构建包含上下文的查询文本
+
+    Args:
+        current_question: 当前问题
+        history_questions: 历史问题列表（按时间正序，最旧的在前）
+
+    Returns:
+        拼接后的查询文本
+    """
+    if not settings.MULTI_TURN_EMBEDDING_ENABLED or not history_questions:
+        return current_question
+
+    max_history = settings.MULTI_TURN_HISTORY_COUNT
+    recent_history = history_questions[-max_history:] if history_questions else []
+
+    if not recent_history:
+        return current_question
+
+    # 拼接：历史问题 + 当前问题
+    context_parts = recent_history + [current_question]
+
+    # 使用分隔符拼接，保持语义连贯
+    context_query = " | ".join(context_parts)
+
+    SQLBotLogUtil.info(f"Context query for embedding: {context_query}")
+
+    return context_query
+
+
+def get_table_embedding(tables: list[dict], question: str, history_questions: List[str] = None):
     _list = []
     for table in tables:
         _list.append({"id": table.get('id'), "schema_table": table.get('schema_table'), "cosine_similarity": 0.0})
@@ -25,7 +57,9 @@ def get_table_embedding(tables: list[dict], question: str):
             end_time = time.time()
             SQLBotLogUtil.info(str(end_time - start_time))
 
-            q_embedding = model.embed_query(question)
+            # 构建包含上下文的查询
+            context_query = build_context_query(question, history_questions)
+            q_embedding = model.embed_query(context_query)
             for index in range(len(results)):
                 item = results[index]
                 _list[index]['cosine_similarity'] = cosine_similarity(q_embedding, item)
@@ -40,7 +74,18 @@ def get_table_embedding(tables: list[dict], question: str):
     return _list
 
 
-def calc_table_embedding(tables: list[dict], question: str):
+def calc_table_embedding(tables: list[dict], question: str, history_questions: List[str] = None):
+    """
+    计算表结构与问题的embedding相似度
+
+    Args:
+        tables: 表结构列表
+        question: 当前问题
+        history_questions: 历史问题列表（可选，用于多轮对话）
+
+    Returns:
+        按相似度排序的表列表
+    """
     _list = []
     for table in tables:
         _list.append(
@@ -58,7 +103,9 @@ def calc_table_embedding(tables: list[dict], question: str):
             # SQLBotLogUtil.info(str(end_time - start_time))
             results = [item.get('embedding') for item in _list]
 
-            q_embedding = model.embed_query(question)
+            # 构建包含上下文的查询
+            context_query = build_context_query(question, history_questions)
+            q_embedding = model.embed_query(context_query)
             for index in range(len(results)):
                 item = results[index]
                 if item: