Update app.py

app.py

@@ -240,19 +240,33 @@ def synthesize_response(state: AgentState) -> AgentState:
     if state.get("answer"):
         print(f"[SYNTHESIZE] skipped — answer already set by tool")
         return state
-    prompt = [
-        SystemMessage(content=get_prompt("final_llm_system")),
+
+    # Pass 1: chain-of-thought reasoning
+    reasoning_prompt = [
+        SystemMessage(content=get_prompt("reasoning_system")),
         HumanMessage(
             content=get_prompt(
-                prompt_key="final_llm_user",
+                prompt_key="reasoning_user",
                 question=state["question"],
                 context=state["context"],
             )
         ),
     ]
-    raw = _llm_answer.invoke(prompt).content.strip()
-    print(f"\n[REASONING]\n{raw}\n")
-    state["answer"] = raw
+    reasoning = _llm_answer.invoke(reasoning_prompt).content.strip()
+    print(f"\n[REASONING]\n{reasoning}\n")
+
+    # Pass 2: extract the final answer cleanly
+    extract_prompt = [
+        SystemMessage(content=get_prompt("extract_system")),
+        HumanMessage(
+            content=get_prompt(
+                prompt_key="extract_user",
+                reasoning=reasoning,
+            )
+        ),
+    ]
+    answer = _llm_answer.invoke(extract_prompt).content.strip()
+    state["answer"] = answer
     return state
 
 def format_output(state: AgentState) -> AgentState: