Skip to content

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

Fix: force set enable_thinking to true for preview qwen3 model (#17604)

Lynn133d9ad07764
diff --git a/rag/advanced_rag/knowlege_compile/_common.py b/rag/advanced_rag/knowlege_compile/_common.py
index 1fceea9408..f426b61071 100644
--- a/rag/advanced_rag/knowlege_compile/_common.py
+++ b/rag/advanced_rag/knowlege_compile/_common.py
@@ -63,7 +63,9 @@ def knowledge_compile_gen_conf(chat_mdl, gen_conf: Optional[dict] = None) -> dic
conf["extra_body"] = extra_body
elif "qwen3" in model_name:
# chat_model.py maps this flag to the provider-specific request body.
- conf["enable_thinking"] = False
+ # -preview variants (e.g. qwen3.8-max-preview) only accept
+ # enable_thinking=True on their API endpoint.
+ conf["enable_thinking"] = True if "-preview" in model_name else False
else:
# LiteLLM maps this common control for providers that support it and
# drops it for providers that do not. Keep model-specific overrides
diff --git a/rag/llm/chat_model.py b/rag/llm/chat_model.py
index c3a917528d..d4015e1673 100644
--- a/rag/llm/chat_model.py
+++ b/rag/llm/chat_model.py
@@ -151,7 +151,12 @@ def _apply_model_family_policies(
# Qwen3 keeps RAGFlow's system default of disabling thinking unless explicitly overridden.
if "qwen3" in model_name_lower:
_pop_thinking_controls()
- enable_thinking = thinking_type == "enabled" if thinking_type else False
+ # -preview variants (e.g. qwen3.8-max-preview) only accept
+ # enable_thinking=True; the API rejects any other value.
+ if "-preview" in model_name_lower:
+ enable_thinking = True
+ else:
+ enable_thinking = thinking_type == "enabled" if thinking_type else False
if backend == "litellm" and provider in {
SupportedLiteLLMProvider.Tongyi_Qianwen,
SupportedLiteLLMProvider.Dashscope,
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 cc9ce65351..ae512abcee 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
@@ -46,6 +46,33 @@ def test_qwen3_can_enable_thinking_explicitly():
assert kwargs["extra_body"] == {"seed": 1, "enable_thinking": True}
+def test_qwen3_preview_variant_forces_thinking_true():
+ """qwen3.x-preview models (e.g. qwen3.8-max-preview) only accept enable_thinking=True."""
+ gen_conf, kwargs = _apply_model_family_policies(
+ "qwen3.8-max-preview",
+ backend="base",
+ gen_conf={},
+ request_kwargs={},
+ )
+
+ assert gen_conf == {}
+ assert kwargs["extra_body"]["enable_thinking"] is True
+
+
+def test_qwen3_preview_ignores_disabled_thinking():
+ """Even with thinking=disabled, -preview still forces enable_thinking=True."""
+ gen_conf, kwargs = _apply_model_family_policies(
+ "qwen3.8-max-preview",
+ backend="base",
+ gen_conf={"thinking": "disabled", "temperature": 0.2},
+ request_kwargs={},
+ )
+
+ assert "thinking" not in gen_conf
+ assert gen_conf == {"temperature": 0.2}
+ assert kwargs["extra_body"]["enable_thinking"] is True
+
+
@pytest.mark.parametrize(
"provider",
[SupportedLiteLLMProvider.Tongyi_Qianwen, SupportedLiteLLMProvider.Dashscope],