diff --git a/backend/apps/chat/task/llm.py b/backend/apps/chat/task/llm.py
index 2ff94400..7aef6812 100644
--- a/backend/apps/chat/task/llm.py
+++ b/backend/apps/chat/task/llm.py
@@ -42,7 +42,8 @@
 from apps.datasource.crud.datasource import get_table_schema, get_tables_sample_data
 from apps.datasource.crud.permission import get_row_permission_filters, is_normal_user
 from apps.datasource.embedding.ds_embedding import get_ds_embedding
-from apps.datasource.models.datasource import CoreDatasource
+from apps.datasource.models.datasource import CoreDatasource, DatasourceConf
+from apps.datasource.utils.utils import aes_decrypt
 from apps.db.db import exec_sql, get_version, check_connection, get_sqlglot_dialect
 from apps.system.crud.aimodel_manage import get_ai_model_list_by_workspace
 from apps.system.crud.assistant import AssistantOutDs, AssistantOutDsFactory, get_assistant_ds
@@ -64,6 +65,59 @@
 
 executor = ThreadPoolExecutor(max_workers=200)
 
+
+# ====================================================================
+# 【改进点1】嵌入态数据源身份翻译：连接指纹映射
+# 嵌入态 self.ds 为外部系统（DataEase）侧 AssistantOutDsSchema，其 id 与
+# SQLBot 侧 core_datasource.id 属不同 id 空间，无法直连。以连接指纹
+# (type, host, port, database, schema) —— 数据源的物理属性 —— 做匹配，
+# 不受两侧人工命名差异影响（名称匹配不可靠：异名同库会漏、同名异库会错命中）。
+# 规则：host 归一（localhost/::1/0.0.0.0 → 127.0.0.1，写法漂移实测存在）；
+#       database 必须参与（同实例多库仅靠 ip+port 不可分）；
+#       SSL 不参与匹配（DataEase→SQLBot 传输 VO 无 SSL 字段，两侧无对齐值）。
+# 映射失败返回 None → 检索回退为仅按高级应用锚（安全侧失效：少过滤而非错过滤）。
+# ====================================================================
+def _normalize_host(host) -> str:
+    """host 归一化：localhost/::1/0.0.0.0 与 127.0.0.1 视为同一地址。"""
+    h = (host or '').strip().lower()
+    return '127.0.0.1' if h in ('localhost', '::1', '0.0.0.0') else h
+
+
+def _ds_fingerprint(ds_type, host, port, database, schema='') -> tuple:
+    """连接指纹：数据源的物理身份五元组（schema 仅 PG 类有值，MySQL 恒为空串）。"""
+    return ((ds_type or '').strip().lower(), _normalize_host(host),
+            int(port or 0), (database or '').strip(), (schema or '').strip())
+
+
+def resolve_embedded_ds_id_by_fingerprint(session: Session, oid: int, out_ds) -> Optional[int]:
+    """嵌入态：按连接指纹把外部系统侧数据源映射到 SQLBot 侧 core_datasource.id。
+    out_ds 为 AssistantOutDsSchema（convert2schema 已解密，host/port/dataBase/db_schema 为明文）。
+    SQLBot 侧 configuration 经 aes_decrypt 解密为 DatasourceConf 后取同构字段比对。
+    映射失败返回 None，检索回退为仅按高级应用锚。"""
+    if out_ds is None:
+        return None
+    ds_type = (getattr(out_ds, 'type', '') or '').strip().lower()
+    if ds_type in ('excel', 'api', 'es'):   # 无物理连接语义的类型不参与指纹
+        return None
+    fp = _ds_fingerprint(ds_type, getattr(out_ds, 'host', None), getattr(out_ds, 'port', None),
+                         getattr(out_ds, 'dataBase', None), getattr(out_ds, 'db_schema', None))
+    try:
+        rows = session.execute(
+            select(CoreDatasource.id, CoreDatasource.type, CoreDatasource.configuration)
+            .where(CoreDatasource.oid == oid)
+        ).all()
+        for ds_id, ds_t, conf_enc in rows:
+            try:
+                conf = DatasourceConf(**json.loads(aes_decrypt(conf_enc)))
+            except Exception:
+                continue    # excel 等类型 configuration 结构不同，跳过
+            if _ds_fingerprint(ds_t, conf.host, conf.port, conf.database, conf.dbSchema) == fp:
+                return ds_id
+        return None
+    except Exception:
+        SQLBotLogUtil.error("resolve_embedded_ds_id_by_fingerprint failed")
+        return None
+
 dynamic_ds_types = [1, 3]
 dynamic_subsql_prefix = 'select * from sqlbot_dynamic_temp_table_'
 
@@ -383,15 +437,15 @@ def filter_terminology_template(self, _session: Session, oid: int = None, ds_id:
         calculate_ds_id = ds_id
         if self.current_assistant:
             calculate_oid = self.current_assistant.oid if self.current_assistant.type != 4 else self.oid
-            if self.current_assistant.type == 1:
-                calculate_ds_id = None
+            # 【改进点3】高级应用态不再丢弃 ds_id：保留用于数据源过滤，应用锚仍并行传递（双锚叠加）
         # 使用原始提取的关键词（未扩展）进行术语匹配，避免同义词重复查询
         match_text = self.chat_question.extracted_keywords or self.chat_question.question
         if self.current_assistant and self.current_assistant.type == 1:
             self.chat_question.terminologies, term_list = get_terminology_template(_session,
                                                                                    match_text,
                                                                                    calculate_oid,
-                                                                                   None, self.current_assistant.id, config=self.embedding_config)
+                                                                                   calculate_ds_id,
+                                                                                   self.current_assistant.id, config=self.embedding_config)
         else:
             self.chat_question.terminologies, term_list = get_terminology_template(_session,
                                                                                    match_text,
@@ -445,13 +499,12 @@ def filter_training_template(self, _session: Session, oid: int = None, ds_id: in
         calculate_ds_id = ds_id
         if self.current_assistant:
             calculate_oid = self.current_assistant.oid if self.current_assistant.type != 4 else self.oid
-            if self.current_assistant.type == 1:
-                calculate_ds_id = None
+            # 【改进点4】高级应用态不再丢弃 ds_id：保留用于数据源过滤，应用锚仍并行传递（双锚叠加）
         if self.current_assistant and self.current_assistant.type == 1:
             self.chat_question.data_training, example_list = get_training_template(_session,
                                                                                    self.chat_question.question,
                                                                                    calculate_oid,
-                                                                                   None, self.current_assistant.id, config=self.embedding_config)
+                                                                                   calculate_ds_id, self.current_assistant.id, config=self.embedding_config)
         else:
             self.chat_question.data_training, example_list = get_training_template(_session,
                                                                                    self.chat_question.question,
@@ -918,6 +971,11 @@ def select_datasource(self, _session: Session):
         if self.ds:
             oid = self.ds.oid if isinstance(self.ds, CoreDatasource) else 1
             ds_id = self.ds.id if isinstance(self.ds, CoreDatasource) else None
+            # 【改进点2】嵌入态：外部系统（DataEase）侧数据源 id 与 SQLBot 侧不同空间，
+            # 按连接指纹 (type, host, port, database, schema) 映射出 SQLBot 侧 ds_id，
+            # 供术语/示例检索做数据源过滤（映射失败回退 None）
+            if ds_id is None:
+                ds_id = resolve_embedded_ds_id_by_fingerprint(_session, oid, self.ds)
 
             # 提取关键词（前置步骤，结果复用于术语匹配和表匹配）
             if self.embedding_config.keyword_enabled:
@@ -1400,6 +1458,11 @@ def run_task(self, in_chat: bool = True, stream: bool = True,
             if self.ds:
                 oid = self.ds.oid if isinstance(self.ds, CoreDatasource) else 1
                 ds_id = self.ds.id if isinstance(self.ds, CoreDatasource) else None
+                # 【改进点2】嵌入态：外部系统（DataEase）侧数据源 id 与 SQLBot 侧不同空间，
+                # 按连接指纹 (type, host, port, database, schema) 映射出 SQLBot 侧 ds_id，
+                # 供术语/示例检索做数据源过滤（映射失败回退 None）
+                if ds_id is None:
+                    ds_id = resolve_embedded_ds_id_by_fingerprint(_session, oid, self.ds)
 
                 # 提取关键词（前置步骤，结果复用于术语匹配和表匹配）
                 if self.embedding_config.keyword_enabled:
diff --git a/backend/apps/data_training/curd/data_training.py b/backend/apps/data_training/curd/data_training.py
index f66c0575..e4d1f4b6 100644
--- a/backend/apps/data_training/curd/data_training.py
+++ b/backend/apps/data_training/curd/data_training.py
@@ -481,6 +481,20 @@ def save_embeddings(session_maker, ids: List[int]):
 LIMIT :top_count
 """
 
+# 【改进点7】高级应用/嵌入态专用：应用锚 + 数据源锚叠加的向量检索（示例侧单值绑定，精确等值）
+embedding_sql_in_advanced_application_and_ds = """
+SELECT id, advanced_application, question, similarity
+FROM
+(SELECT id, advanced_application, datasource, question, oid, enabled,
+( 1 - (embedding <=> :embedding_array) ) AS similarity
+FROM data_training AS child
+) TEMP
+WHERE similarity > :similarity and oid = :oid and advanced_application = :advanced_application and enabled = true
+and datasource = :datasource
+ORDER BY similarity DESC
+LIMIT :top_count
+"""
+
 
 def select_training_by_question(session: SessionDep, question: str, oid: int, datasource: Optional[int] = None,
                                 advanced_application_id: Optional[int] = None, *, config: EmbeddingConfig | None = None):
@@ -504,6 +518,9 @@ def select_training_by_question(session: SessionDep, question: str, oid: int, da
     )
     if advanced_application_id is not None:
         stmt = stmt.where(and_(DataTraining.advanced_application == advanced_application_id))
+        # 【改进点8】高级应用/嵌入态叠加数据源过滤（示例侧单值绑定，精确等值）
+        if datasource is not None:
+            stmt = stmt.where(DataTraining.datasource == datasource)
     else:
         stmt = stmt.where(and_(DataTraining.datasource == datasource))
 
@@ -519,7 +536,13 @@ def select_training_by_question(session: SessionDep, question: str, oid: int, da
 
                 embedding = parse_valid_embedding(model.embed_query(question))
 
-                if advanced_application_id is not None:
+                if advanced_application_id is not None and datasource is not None:
+                    # 【改进点8】双锚齐备：应用锚 + 数据源过滤叠加
+                    results = session.execute(text(embedding_sql_in_advanced_application_and_ds),
+                                              {'embedding_array': str(embedding), 'oid': oid, 'similarity': config.data_training_similarity, 'top_count': config.data_training_top_count,
+                                               'advanced_application': advanced_application_id,
+                                               'datasource': datasource})
+                elif advanced_application_id is not None:
                     results = session.execute(text(embedding_sql_in_advanced_application),
                                               {'embedding_array': str(embedding), 'oid': oid, 'similarity': config.data_training_similarity, 'top_count': config.data_training_top_count,
                                                'advanced_application': advanced_application_id})
diff --git a/backend/apps/terminology/curd/terminology.py b/backend/apps/terminology/curd/terminology.py
index 0cb50548..3de101d5 100644
--- a/backend/apps/terminology/curd/terminology.py
+++ b/backend/apps/terminology/curd/terminology.py
@@ -801,6 +801,26 @@ def save_embeddings(session_maker, ids: List[int]):
 LIMIT :top_count
 """
 
+# 【改进点5】高级应用/嵌入态专用：应用锚 + 数据源锚叠加的向量检索
+# 语义：命中应用锚，且（所有数据源生效 或 绑定数据源数组包含当前所选 ds）
+embedding_sql_with_advanced_application_and_ds = """
+SELECT id, pid, word, similarity
+FROM
+(SELECT id, pid, word, oid, specific_ds, advanced_application, datasource_ids, enabled,
+( 1 - (embedding <=> :embedding_array) ) AS similarity
+FROM terminology AS child
+) TEMP
+WHERE similarity > :similarity AND oid = :oid AND enabled = true
+AND advanced_application = :advanced_application_id
+AND (
+    (specific_ds = false OR specific_ds IS NULL)
+    OR
+    (specific_ds = true AND datasource_ids IS NOT NULL AND datasource_ids @> jsonb_build_array(:datasource))
+)
+ORDER BY similarity DESC
+LIMIT :top_count
+"""
+
 
 def select_terminology_by_word(session: SessionDep, word: str, oid: int, datasource: int = None,
                                advanced_application_id: Optional[int] = None, *, config: EmbeddingConfig | None = None):
@@ -823,6 +843,18 @@ def select_terminology_by_word(session: SessionDep, word: str, oid: int, datasou
 
     if advanced_application_id is not None:
         stmt = stmt.where(Terminology.advanced_application == advanced_application_id)
+        # 【改进点6】高级应用/嵌入态叠加数据源过滤：放行"所有数据源"，过滤"绑定不含当前 ds"的术语
+        if datasource is not None:
+            stmt = stmt.where(
+                or_(
+                    or_(Terminology.specific_ds == False, Terminology.specific_ds.is_(None)),
+                    and_(
+                        Terminology.specific_ds == True,
+                        Terminology.datasource_ids.isnot(None),
+                        text("datasource_ids @> jsonb_build_array(:datasource)")
+                    )
+                )
+            )
     elif datasource is not None:
         stmt = stmt.where(
             or_(
@@ -854,7 +886,13 @@ def select_terminology_by_word(session: SessionDep, word: str, oid: int, datasou
 
                 embedding = parse_valid_embedding(model.embed_query(word))
 
-                if advanced_application_id is not None:
+                if advanced_application_id is not None and datasource is not None:
+                    # 【改进点6】双锚齐备：应用锚 + 数据源过滤叠加
+                    results = session.execute(text(embedding_sql_with_advanced_application_and_ds),
+                                              {'embedding_array': str(embedding), 'oid': oid, 'similarity': config.terminology_similarity, 'top_count': config.terminology_top_count,
+                                               'advanced_application_id': advanced_application_id,
+                                               'datasource': datasource}).fetchall()
+                elif advanced_application_id is not None:
                     results = session.execute(text(embedding_sql_with_advanced_application),
                                               {'embedding_array': str(embedding), 'oid': oid, 'similarity': config.terminology_similarity, 'top_count': config.terminology_top_count,
                                                'advanced_application_id': advanced_application_id}).fetchall()