Skip to content

RAGFlow is an agent-native database: a Retrieval-Augmented Generation (RAG) engine that grew into context infrastructure.

fix(llm): disable DeepSeek V4 thinking by default via extra_body (#18241)

D819c466ff3b1a4a
diff --git a/agent/canvas.py b/agent/canvas.py
index 14be6c10a6..904d901a7f 100644
--- a/agent/canvas.py
+++ b/agent/canvas.py
@@ -349,7 +349,7 @@ class Canvas(Graph):
"sys.conversation_turns": 0,
"sys.files": [],
"sys.history": [],
- "sys.date": datetime.datetime.now(datetime.timezone.utc).strftime("%Y-%m-%d %H:%M:%S"),
+ "sys.date": datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
}
self.variables = {}
# Aggregated provider token usage (prompt/completion/total) across every LLM
@@ -368,7 +368,7 @@ class Canvas(Graph):
if "sys.history" not in self.globals:
self.globals["sys.history"] = []
if "sys.date" not in self.globals:
- self.globals["sys.date"] = datetime.datetime.now(datetime.timezone.utc).strftime("%Y-%m-%d %H:%M:%S")
+ self.globals["sys.date"] = datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")
else:
self.globals = {
"sys.query": "",
@@ -376,7 +376,7 @@ class Canvas(Graph):
"sys.conversation_turns": 0,
"sys.files": [],
"sys.history": [],
- "sys.date": datetime.datetime.now(datetime.timezone.utc).strftime("%Y-%m-%d %H:%M:%S"),
+ "sys.date": datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
}
if "variables" in self.dsl:
self.variables = self.dsl["variables"]
@@ -482,7 +482,7 @@ class Canvas(Graph):
reset_llm_request_context(_req_ctx_token)
async def _run_impl(self, **kwargs):
- self.globals["sys.date"] = datetime.datetime.now(datetime.timezone.utc).strftime("%Y-%m-%d %H:%M:%S")
+ self.globals["sys.date"] = datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")
st = time.perf_counter()
self._loop = asyncio.get_running_loop()
self.message_id = get_uuid()
diff --git a/internal/agent/runtime/state.go b/internal/agent/runtime/state.go
index f1f9dc1c03..d337fb5aae 100644
--- a/internal/agent/runtime/state.go
+++ b/internal/agent/runtime/state.go
@@ -95,7 +95,7 @@ func NewCanvasState(runID, sessionID string) *CanvasState {
return s
}
-// EnsureSysDate fills sys.date with the current UTC timestamp when it
+// EnsureSysDate fills sys.date with the current local timestamp when it
// is missing or blank. Python canvas initializes the same variable with
// "%Y-%m-%d %H:%M:%S"; keep that wire format for DSL compatibility.
func (s *CanvasState) EnsureSysDate() {
@@ -110,7 +110,7 @@ func (s *CanvasState) EnsureSysDate() {
if v, ok := s.Sys["date"]; ok && strings.TrimSpace(fmt.Sprint(v)) != "" {
return
}
- s.Sys["date"] = time.Now().UTC().Format("2006-01-02 15:04:05")
+ s.Sys["date"] = time.Now().Format("2006-01-02 15:04:05")
}
// init registers CanvasState with eino's internal type registry so
diff --git a/rag/llm/chat_model.py b/rag/llm/chat_model.py
index 577b9a84a9..e5fcbcb7cd 100644
--- a/rag/llm/chat_model.py
+++ b/rag/llm/chat_model.py
@@ -171,7 +171,12 @@ def _apply_model_family_policies(
return sanitized_gen_conf, sanitized_kwargs
if backend == "litellm":
- if provider in {SupportedLiteLLMProvider.OpenAI, SupportedLiteLLMProvider.Azure_OpenAI} and "gpt-5" in model_name_lower:
+ if provider == SupportedLiteLLMProvider.DeepSeek:
+ _pop_thinking_controls()
+ sanitized_gen_conf.pop("reasoning_effort", None)
+ sanitized_kwargs.pop("reasoning_effort", None)
+ _merge_extra_body(sanitized_gen_conf, {"thinking": {"type": thinking_type or "disabled"}})
+ elif provider in {SupportedLiteLLMProvider.OpenAI, SupportedLiteLLMProvider.Azure_OpenAI} and "gpt-5" in model_name_lower:
for key in ("temperature", "top_p", "logprobs", "top_logprobs"):
sanitized_gen_conf.pop(key, None)
sanitized_kwargs.pop(key, None)
diff --git a/test/unit_test/rag/llm/test_chat_model_thinking_policy.py b/test/unit_test/rag/llm/test_chat_model_thinking_policy.py
index 32ea723905..4318fc05ae 100644
--- a/test/unit_test/rag/llm/test_chat_model_thinking_policy.py
+++ b/test/unit_test/rag/llm/test_chat_model_thinking_policy.py
@@ -188,6 +188,43 @@ def test_glm_thinking_maps_to_zhipu_payload():
assert gen_conf["thinking"] == {"type": "enabled"}
+def test_deepseek_thinking_disabled_via_extra_body():
+ # litellm 1.82.x DeepSeek transformation drops `thinking` from the top
+ # level (only {"type": "enabled"} survives) and rejects reasoning_effort
+ # alongside thinking: disabled. The toggle must be carried in extra_body
+ # and any reasoning_effort stripped. When unspecified, default to disabled.
+ for gen_input in (
+ {"thinking": "disabled", "reasoning_effort": "high"},
+ {"thinking": "default"},
+ {"temperature": 0.5},
+ ):
+ gen_conf, kwargs = _apply_model_family_policies(
+ "deepseek-v4-flash",
+ backend="litellm",
+ provider=SupportedLiteLLMProvider.DeepSeek,
+ gen_conf=dict(gen_input),
+ request_kwargs={},
+ )
+ assert kwargs == {}
+ assert "thinking" not in gen_conf
+ assert "reasoning_effort" not in gen_conf
+ assert gen_conf["extra_body"]["thinking"] == {"type": "disabled"}
+
+
+def test_deepseek_thinking_enabled_via_extra_body():
+ gen_conf, kwargs = _apply_model_family_policies(
+ "deepseek-v4-flash",
+ backend="litellm",
+ provider=SupportedLiteLLMProvider.DeepSeek,
+ gen_conf={"thinking": "enabled"},
+ request_kwargs={},
+ )
+
+ assert kwargs == {}
+ assert "thinking" not in gen_conf
+ assert gen_conf["extra_body"]["thinking"] == {"type": "enabled"}
+
+
def test_litellm_provider_body_fields_move_to_extra_body_before_drop_params():
completion_args = {
"model": "kimi-latest",