Update app.py

app.py

@@ -29,34 +29,54 @@ agent = create_react_agent(model=llm, tools=tool_node)
 
 def plan_node(state: AgentState) -> AgentState:
     """
-    Decide which tool (if any) to call. Append the user_input to state["messages"] and
-    return a partial AgentState that sets exactly one of:
-      - web_search_query (string)
-      - ocr_path (string)
-      - excel_path (string) + excel_sheet_name (optional)
-      - final_answer (string)
+    Decide which tool (if any) to call. The state is expected to contain:
+      - state["messages"]: a list of BaseMessage (SystemMessage, HumanMessage, etc.)
+      - state["user_input"]: the raw user question (string)
+
+    We append a new HumanMessage(user_input) to messages, then ask the LLM
+    (via ChatOpenAI) to return exactly one key: web_search_query, ocr_path,
+    excel_path (with excel_sheet_name), or final_answer. The LLM must reply
+    with a bare Python‐dict literal.
+
+    We then return a new partial AgentState that always includes an updated
+    "messages" list plus exactly one of those tool‐request keys (or final_answer).
     """
+    # 1) Pull user_input out of state
     user_input = state.get("user_input", "")
-    prior = state.get("messages", [])
-    chat_history = prior + [f"USER: {user_input}"]
-
-    prompt = chat_history + [
-        "ASSISTANT: You can set exactly one of the following keys in a Python dict:\n"
-        "  • web_search_query: <search terms>  \n"
-        "  • ocr_path: <path to an image>  \n"
-        "  • excel_path: <path to xlsx>  \n"
-        "  • excel_sheet_name: <sheet name>  \n"
-        "Or, if no tool is needed, set final_answer: <your answer>.\n"
-        "Example: {'web_search_query':'Mercedes Sosa discography'}\n"
-        "Respond with only that Python dict literal—no additional text."
-    ]
-    print("plan_node  content problem: ", prompt)
-    llm_out = llm(prompt).strip()
-    print("plan_node  passed")
+    # 2) Grab prior chat history, which should already be a list of BaseMessage
+    prior_msgs = state.get("messages", [])
+    # 3) Append the new user message as a HumanMessage
+    new_history = prior_msgs + [HumanMessage(content=user_input)]
+
+    # 4) Build a prompt that explains how to choose exactly one key
+    #    We leave new_history as a list of BaseMessage; LLM expects that format.
+    explanation = SystemMessage(
+        content=(
+            "You can set exactly one of the following keys (in a Python dict) and nothing else:\n"
+            "  • web_search_query: <search terms>  \n"
+            "  • ocr_path: <path to an image file>  \n"
+            "  • excel_path: <path to a .xlsx file>  \n"
+            "  • excel_sheet_name: <sheet name>  \n"
+            "Or, if no tool is needed, set final_answer: <your answer>.\n"
+            "Example: {'web_search_query':'Mercedes Sosa discography'}\n"
+            "Respond with only that Python dict literal—no extra text or explanation."
+        )
+    )
+
+    # 5) Combine the user conversation with our explanation
+    prompt_messages = new_history + [explanation]
+
+    # 6) Call the LLM. Because prompt_messages is a list of BaseMessage,
+    #    ChatOpenAI will return an AIMessage.
+    llm_response = llm(prompt_messages)
+    llm_out = llm_response.content.strip()
+
+    # 7) Try to eval the LLM response as a Python dict
     try:
         parsed = eval(llm_out, {}, {})
         if isinstance(parsed, dict):
-            new_state: AgentState = {"messages": chat_history}
+            # Build a new state: keep our updated messages, plus exactly one key
+            new_state: AgentState = {"messages": new_history}
             allowed = {
                 "web_search_query",
                 "ocr_path",
@@ -71,9 +91,9 @@ def plan_node(state: AgentState) -> AgentState:
     except Exception:
         pass
 
-    # Fallback if parsing failed
+    # 8) Fallback if parsing failed: keep messages, set a generic final_answer
     return {
-        "messages": chat_history,
+        "messages": new_history,
         "final_answer": "Sorry, I could not parse your intent."
     }