fix(chat_db_qa): tell LLM the table list is a TOP-K subset, not the full DB (#3045)

Co-authored-by: hankchrome <hankchrome@github.com>

packages/dbgpt-app/src/dbgpt_app/scene/chat_db/professional_qa/prompt.py

@@ -18,6 +18,16 @@ available to answer this question." Feel free to fudge information.
 Use the following tables generate sql if have any table info:
 {table_info}
 
+NOTE: this is a QA-only scene; you cannot execute SQL here. The table list \
+above contains only the TOP-K most relevant tables retrieved from a vector \
+store; it is NOT the complete list of tables in the database. If the user \
+asks for a total count of tables, a list of all tables, a schema overview, \
+or any other metadata that requires knowing the whole database, do NOT \
+answer with a count or list derived from the partial table list above. \
+Instead, acknowledge the limitation and show the SQL query against \
+INFORMATION_SCHEMA (or the dialect-specific system catalog) that the user \
+can run in a SQL-executing scene to obtain the answer.
+
 user question:
 {input}
 think step by step.
@@ -27,9 +37,16 @@ _DEFAULT_TEMPLATE_ZH = """
 根据要求和问题，提供专业的答案。如果无法从提供的内容中获取答案，请说：\
 “知识库中提供的信息不足以回答此问题。” 禁止随意捏造信息。
 
-使用以下表结构信息: 
+使用以下表结构信息:
 {table_info}
 
+注意：这是只问答（QA）场景，无法在此处执行 SQL。以上表清单只是从向量库\
+检索到的 TOP-K 最相关表，并非数据库中所有表的完整清单。当用户询问表的\
+总数、所有表的列表、schema 结构总览，或其他需要了解整个数据库元信息的\
+问题时，请勿根据以上部分表清单直接给出数量或列表。请说明此限制，并展示\
+可在「可执行 SQL」场景中运行、针对 INFORMATION_SCHEMA（或方言对应的系统\
+目录）的 SQL 查询，以便用户取得正确答案。
+
 问题:
 {input}
 一步步思考。