fix: vLLM chat_template_kwargs breaks LangChain structured output

vLLM's chat_template_kwargs leaked into LangChain's AsyncCompletions.parse()
via _get_langchain_model's model_kwargs, causing structured decomposition
to fail on vLLM backends. Skip vLLM-specific params when building the
LangChain model — only provider-agnostic params (OpenAI reasoning) pass through.

backend/app/services/llm_client.py

@@ -110,10 +110,13 @@ class LLMClient:
             os.environ.setdefault("OPENAI_API_KEY", self.settings.llm_api_key)
             os.environ.setdefault("OPENAI_BASE_URL", self.settings.llm_base_url)
 
-            # Pass thinking/reasoning disable params via model_kwargs.
+            # vLLM's chat_template_kwargs is incompatible with LangChain's
-            # LangChain's ChatOpenAI forwards model_kwargs as top-level
+            # with_structured_output() — it leaks into AsyncCompletions.parse()
-            # request parameters, which is equivalent to OpenAI SDK's extra_body.
+            # which rejects the unknown kwarg. Only provider-agnostic params
-            model_kwargs = self._build_extra_body() or None
+            # (e.g. OpenAI's reasoning) are safe to pass via model_kwargs.
+            model_kwargs: dict[str, Any] | None = None
+            if not self.settings.vllm_engine and not self.enable_thinking:
+                model_kwargs = {"reasoning": {"enabled": False}}
 
             self._langchain_model = init_chat_model(
                 model=self.model,

backend/requirements.txt

@@ -1,18 +1,18 @@
 fastapi==0.109.0
 uvicorn[standard]==0.27.0
-pydantic==2.5.3
+pydantic>=2.7.4,<3.0.0
-pydantic-settings==2.1.0
+pydantic-settings>=2.1.0
 chromadb==0.4.22
 numpy<2.0
-python-docx==1.1.0
+python-docx>=1.1.0
-pypdf==4.0.2
+pypdf>=4.0.2
-python-dotenv==1.0.0
+python-dotenv>=1.0.0
-httpx==0.26.0
+httpx>=0.26.0
-openai==1.12.0
+openai>=2.26.0,<3.0.0
 pytest==7.4.4
 pytest-asyncio==0.23.4
-tiktoken==0.5.2
+tiktoken>=0.7.0,<1.0.0
-python-multipart==0.0.6
+python-multipart>=0.0.6
-reportlab==4.2.5
+reportlab>=4.2.5
-langchain==1.2.12
+langchain>=1.2.12,<1.3.0
-langchain-openai==1.1.11
+langchain-openai>=1.1.11,<1.2.0