Update app.py

app.py

@@ -91,39 +91,30 @@ def plan_node(state: AgentState) -> AgentState:
 
 # ─── 3) Revised finalize_node ───
 def finalize_node(state: AgentState) -> AgentState:
-    """
-    Collect any tool results from state and then ask the LLM for a final answer.
-    We build a fresh list of SystemMessages for tool results (no reuse of prior AIMessage).
-    """
-    # 1) Create a list of SystemMessages for each available tool result
-    messages_for_llm = []
-    if state.get("web_search_result") is not None:
-        messages_for_llm.append(
-            SystemMessage(content=f"WEB_SEARCH_RESULT: {state['web_search_result']}")
-        )
-    if state.get("ocr_result") is not None:
-        messages_for_llm.append(
-            SystemMessage(content=f"OCR_RESULT: {state['ocr_result']}")
-        )
-    if state.get("excel_result") is not None:
-        messages_for_llm.append(
-            SystemMessage(content=f"EXCEL_RESULT: {state['excel_result']}")
-        )
-
-    # 2) If plan_node already set final_answer, return it without calling LLM again
+    # If plan_node already provided a final answer, skip LLM
     if state.get("final_answer") is not None:
         return {"final_answer": state["final_answer"]}
 
-    # 3) Otherwise, append our "please give final answer" SystemMessage
-    messages_for_llm.append(
-        SystemMessage(content="Please provide the final answer now.")
-    )
+    # Re-extract the last user question from messages
+    question = ""
+    for msg in reversed(state.get("messages", [])):
+        if isinstance(msg, HumanMessage):
+            question = msg.content
+            break
 
-    # 4) Call the LLM with our fresh list of SystemMessages
-    llm_response = llm(messages_for_llm)
+    # Build a combined context
+    combined = f"USER_QUESTION: {question}\n"
+    if sr := state.get("web_search_result"):
+        combined += f"WEB_SEARCH_RESULT: {sr}\n"
+    if orc := state.get("ocr_result"):
+        combined += f"OCR_RESULT: {orc}\n"
+    if exr := state.get("excel_result"):
+        combined += f"EXCEL_RESULT: {exr}\n"
+    combined += "Based on the above, provide the final answer."
+
+    llm_response = llm([SystemMessage(content=combined)])
     return {"final_answer": llm_response.content.strip()}
 
-
 # ─── 4) Wrap tools in a ToolNode ───
 tool_node = ToolNode([web_search_tool, ocr_image_tool, parse_excel_tool])
 
@@ -185,19 +176,15 @@ def respond_to_input(user_input: str) -> str:
     Then invoke the graph; return the final_answer from the resulting state.
     """
     system_msg = SystemMessage(
-        content=(
-            "You have access to exactly these tools:\n"
-            "  1) web_search(query:str) → Returns DuckDuckGo results.\n"
-            "  2) parse_excel(path:str, sheet_name:str) → Reads an Excel file.\n"
-            "  3) ocr_image(path:str) → Runs OCR on an image.\n\n"
-            "If you need a tool, set exactly one of these keys in a Python dict:\n"
-            "  • web_search_query: <search terms>\n"
-            "  • ocr_path: <path to image>\n"
-            "  • excel_path: <path to xlsx>\n"
-            "  • excel_sheet_name: <sheet name>\n"
-            "Otherwise, set final_answer: <your answer>.\n"
-            "Respond with only that Python dict literal—no extra text."
-        )
+    content=(
+        "You are an agent that decides whether to call a tool or answer the user directly. "
+        "The user’s question is below. If the answer can be given directly, return {'final_answer': <your answer>}."
+        "If you need to call a tool, set exactly one key from the following in a Python dict: "
+        "  • web_search_query: <search terms>\n"
+        "  • ocr_path: <path to an image file>\n"
+        "  • excel_path: <path to a .xlsx file>, excel_sheet_name: <sheet name>.\n"
+        "Do not include any extra text or markdown—only return a valid Python dict literal."
+    )
     )
     human_msg = HumanMessage(content=user_input)