diff --git a/api/db/services/llm_service.py b/api/db/services/llm_service.py
index 2a41918be9..148cba7373 100644
--- a/api/db/services/llm_service.py
+++ b/api/db/services/llm_service.py
@@ -86,6 +86,18 @@ class LLMBundle(LLM4Tenant):
         """Release resources held by this LLMBundle instance."""
         super().close()
 
+    def clone(self):
+        kwargs = {
+            "trace_context": dict(self.trace_context or {}),
+            "langfuse_session_id": self.langfuse_session_id,
+            "verbose_tool_use": self.verbose_tool_use,
+        }
+        for attr, key in (("max_retries", "max_retries"), ("base_delay", "retry_interval"), ("max_rounds", "max_rounds")):
+            value = getattr(self.mdl, attr, None)
+            if value is not None:
+                kwargs[key] = value
+        return LLMBundle(self.tenant_id, dict(self.model_config), lang=getattr(self, "lang", "Chinese"), **kwargs)
+
     def __enter__(self):
         """Enter context manager."""
         return self
diff --git a/api/db/services/tenant_llm_service.py b/api/db/services/tenant_llm_service.py
index 511ee1a950..efe47f1a6d 100644
--- a/api/db/services/tenant_llm_service.py
+++ b/api/db/services/tenant_llm_service.py
@@ -508,6 +508,7 @@ class LLM4Tenant:
         self.trace_context = kwargs.pop("trace_context", None) or {}
         self.langfuse_session_id = kwargs.pop("langfuse_session_id", None)
         self.tenant_id = tenant_id
+        self.lang = lang
         self.llm_name = model_config["llm_name"]
         self.model_config = model_config
         self.mdl = TenantLLMService.model_instance(model_config, lang=lang, **kwargs)
diff --git a/rag/advanced_rag/agentic_rag.py b/rag/advanced_rag/agentic_rag.py
index ac617d1c7e..32d0db729b 100644
--- a/rag/advanced_rag/agentic_rag.py
+++ b/rag/advanced_rag/agentic_rag.py
@@ -82,7 +82,7 @@ class RAGTools:
         thinking_mode: str = "medium",
     ):
         self.tenant_ids = tenant_ids
-        self.chat_mdl = chat_mdl
+        self.chat_mdl = chat_mdl.clone()
         self.embed_mdl = embed_mdl
         self.thinking_mode = thinking_mode
         self.field_map = {}
diff --git a/rag/advanced_rag/harness/agent.py b/rag/advanced_rag/harness/agent.py
index 1fe2223b5a..c12a6ab286 100644
--- a/rag/advanced_rag/harness/agent.py
+++ b/rag/advanced_rag/harness/agent.py
@@ -1,6 +1,6 @@
 """Research Agent — inner tool-calling loop for high/ultra modes.
 
-Native tool-calling: a chat model deep-copied from ``tools.chat_mdl`` is bound
+Native tool-calling: a chat model cloned from ``tools.chat_mdl`` is bound
 (via ``bind_tools``) to the phase-gated tool schemas plus ``think_tool`` /
 ``generate_report``, and a lightweight session routes each tool call to the
 harness pipeline. Binding onto a *copy* keeps the shared ``tools.chat_mdl``
@@ -14,7 +14,6 @@ that the loop parses.
 import json
 import logging
 import re
-from copy import deepcopy
 
 from rag.advanced_rag.harness.types import ClaimTarget, ExecutionStrategy, ToolResult
 from rag.advanced_rag.harness.pipeline import Pipeline
@@ -158,8 +157,8 @@ async def research_agent_loop(
         has_routed_scope=bool(getattr(pipeline, "_routed_docs", None)),
     )
 
-    # Deep-copy so binding tools never leaks onto the shared chat model.
-    agent_mdl = deepcopy(tools.chat_mdl)
+    # Clone so binding tools never leaks onto the shared chat model.
+    agent_mdl = tools.chat_mdl.clone()
     if getattr(agent_mdl, "is_tools", False):
         return await _research_native(claim, agent_mdl, pipeline, phase, phase_config, gated_defs, mode)