Add zai-sdk dependency and update model references to glm-4.5-air

backend/examples/cli_research.py

@@ -21,7 +21,7 @@ def main() -> None:
     )
     parser.add_argument(
         "--reasoning-model",
-        default="GLM-4.7",
+        default="glm-4.5-air",
         help="Model for the final answer",
     )
     args = parser.parse_args()

backend/pyproject.toml

@@ -18,6 +18,7 @@ dependencies = [
     "langgraph-api",
     "fastapi",
     "tavily-python",
+    "zai-sdk",
 ]
 
 

backend/src/agent/configuration.py

@@ -9,21 +9,21 @@ class Configuration(BaseModel):
     """The configuration for the agent."""
 
     query_generator_model: str = Field(
-        default="GLM-4.7",
+        default="glm-4.5-air",
         metadata={
             "description": "The name of the language model to use for the agent's query generation."
         },
     )
 
     reflection_model: str = Field(
-        default="GLM-4.7",
+        default="glm-4.5-air",
         metadata={
             "description": "The name of the language model to use for the agent's reflection."
         },
     )
 
     answer_model: str = Field(
-        default="GLM-4.7",
+        default="glm-4.5-air",
         metadata={
             "description": "The name of the language model to use for the agent's answer."
         },

backend/src/agent/graph.py

@@ -11,6 +11,7 @@ from langgraph.graph import START, END
 from langchain_core.runnables import RunnableConfig
 from langchain_openai import ChatOpenAI
 from tavily import TavilyClient
+from zai import ZhipuAiClient
 
 from agent.state import (
     OverallState,
@@ -40,6 +41,7 @@ if os.getenv("TAVILY_API_KEY") is None:
     raise ValueError("TAVILY_API_KEY is not set")
 
 tavily_client = TavilyClient(api_key=os.getenv("TAVILY_API_KEY"))
+zhipu_client = ZhipuAiClient(api_key=os.getenv("OPENAI_API_KEY"))
 SchemaT = TypeVar("SchemaT", bound=BaseModel)
 
 
@@ -73,13 +75,17 @@ def _extract_json_payload(content: str) -> str:
 
 
 def _invoke_json_model(
-    llm: ChatOpenAI,
+    model: str,
     prompt: str,
     schema: type[SchemaT],
 ) -> SchemaT:
-    """Invoke the model for JSON text and validate it with Pydantic."""
-    result = llm.invoke(prompt)
-    content = result.content if isinstance(result.content, str) else str(result.content)
+    """Invoke the model via zai-sdk JSON mode and validate with Pydantic."""
+    response = zhipu_client.chat.completions.create(
+        model=model,
+        messages=[{"role": "user", "content": prompt}],
+        response_format={"type": "json_object"},
+    )
+    content = response.choices[0].message.content
     json_payload = _extract_json_payload(content)
 
     try:
@@ -93,22 +99,25 @@ def _invoke_json_model(
 def _supports_native_structured_output(model: str) -> bool:
     """Return whether the selected model should use LangChain structured output."""
     normalized = model.strip().lower()
-    return normalized.startswith("glm-4.7")
+    return not normalized.startswith("glm")
 
 
 def _invoke_structured_model(
-    llm: ChatOpenAI,
     model: str,
     prompt: str,
     schema: type[SchemaT],
 ) -> SchemaT:
-    """Use native structured output when supported, otherwise fall back to JSON parsing."""
+    """Use direct provider JSON mode for GLM, otherwise try native structured output."""
+    if not _supports_native_structured_output(model):
+        return _invoke_json_model(model, prompt, schema)
+
     if _supports_native_structured_output(model):
         try:
+            llm = create_llm(model, temperature=1.0)
             return llm.with_structured_output(schema).invoke(prompt)
         except ValueError:
-            return _invoke_json_model(llm, prompt, schema)
-    return _invoke_json_model(llm, prompt, schema)
+            return _invoke_json_model(model, prompt, schema)
+    return _invoke_json_model(model, prompt, schema)
 
 
 # Nodes
@@ -130,8 +139,6 @@ def generate_query(state: OverallState, config: RunnableConfig) -> QueryGenerati
     if state.get("initial_search_query_count") is None:
         state["initial_search_query_count"] = configurable.number_of_initial_queries
 
-    llm = create_llm(configurable.query_generator_model, temperature=1.0)
-
     current_date = get_current_date()
     formatted_prompt = query_writer_instructions.format(
         current_date=current_date,
@@ -139,7 +146,6 @@ def generate_query(state: OverallState, config: RunnableConfig) -> QueryGenerati
         number_queries=state["initial_search_query_count"],
     )
     result = _invoke_structured_model(
-        llm,
         configurable.query_generator_model,
         formatted_prompt,
         SearchQueryList,
@@ -216,9 +222,7 @@ def reflection(state: OverallState, config: RunnableConfig) -> ReflectionState:
         research_topic=get_research_topic(state["messages"]),
         summaries="\n\n---\n\n".join(state["web_research_result"]),
     )
-    llm = create_llm(reasoning_model, temperature=1.0)
     result = _invoke_structured_model(
-        llm,
         reasoning_model,
         formatted_prompt,
         Reflection,

backend/src/agent/prompts.py

@@ -25,12 +25,10 @@ Format:
 Example:
 
 Topic: What revenue grew more last year apple stock or the number of people buying an iphone
-```json
 {{
     "rationale": "To answer this comparative growth question accurately, we need specific data points on Apple's stock performance and iPhone sales metrics. These queries target the precise financial information needed: company revenue trends, product-specific unit sales figures, and stock price movement over the same fiscal period for direct comparison.",
     "query": ["Apple total revenue growth fiscal year 2024", "iPhone unit sales growth fiscal year 2024", "Apple stock price growth fiscal year 2024"]
 }}
-```
 
 Context: {research_topic}"""
 
@@ -67,13 +65,11 @@ Output Format:
    - "follow_up_queries": Write a specific question to address this gap
 
 Example:
-```json
 {{
-    "is_sufficient": true, // or false
-    "knowledge_gap": "The summary lacks information about performance metrics and benchmarks", // "" if is_sufficient is true
-    "follow_up_queries": ["What are typical performance benchmarks and metrics used to evaluate [specific technology]?"] // [] if is_sufficient is true
+    "is_sufficient": true,
+    "knowledge_gap": "",
+    "follow_up_queries": []
 }}
-```
 
 Reflect carefully on the Summaries to identify knowledge gaps and produce a follow-up query. Then, produce your output following this JSON format:
 

frontend/src/components/InputForm.tsx

@@ -26,7 +26,7 @@ export const InputForm: React.FC<InputFormProps> = ({
 }) => {
   const [internalInputValue, setInternalInputValue] = useState("");
   const [effort, setEffort] = useState("medium");
-  const [model, setModel] = useState("GLM-4.7");
+  const [model, setModel] = useState("glm-4.5-air");
 
   const handleInternalSubmit = (e?: React.FormEvent) => {
     if (e) e.preventDefault();
@@ -135,6 +135,14 @@ export const InputForm: React.FC<InputFormProps> = ({
                 <SelectValue placeholder="Model" />
               </SelectTrigger>
               <SelectContent className="bg-neutral-700 border-neutral-600 text-neutral-300 cursor-pointer">
+                <SelectItem
+                  value="glm-4.5-air"
+                  className="hover:bg-neutral-600 focus:bg-neutral-600 cursor-pointer"
+                >
+                  <div className="flex items-center">
+                    <Zap className="h-4 w-4 mr-2 text-emerald-400" /> glm-4.5-air
+                  </div>
+                </SelectItem>
                 <SelectItem
                   value="GLM-4.7"
                   className="hover:bg-neutral-600 focus:bg-neutral-600 cursor-pointer"

frontend/src/components/WelcomeScreen.tsx

@@ -33,7 +33,7 @@ export const WelcomeScreen: React.FC<WelcomeScreenProps> = ({
       />
     </div>
     <p className="text-xs text-neutral-500">
-      Powered by GLM-4.7, Tavily, and LangChain LangGraph.
+      Powered by glm-4.5-air, GLM-4.7, Tavily, and LangChain LangGraph.
     </p>
   </div>
 );