Update app.py

app.py

@@ -28,12 +28,6 @@ llm = ChatOpenAI(model_name="gpt-4.1-mini", temperature=0.0)
 
 # ─── 2) Revised plan_node ───
 def plan_node(state: AgentState) -> AgentState:
-    """
-    Look at the last HumanMessage in state['messages'] to get user_input.
-    Then call llm with exactly [SystemMessage, HumanMessage(user_input)] so
-    we never feed in a list lacking an AIMessage internally.
-    """
-    # 1) Find the last HumanMessage from prior history
     prior_msgs = state.get("messages", [])
     user_input = ""
     for msg in reversed(prior_msgs):
@@ -41,30 +35,39 @@ def plan_node(state: AgentState) -> AgentState:
             user_input = msg.content
             break
 
-    # 2) Build a fresh SystemMessage explaining exactly one dict key
+    # (1) Build your system/human messages exactly as before
     system_msg = SystemMessage(
         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"
-            "  • audio_path: <path to an audio file>\n"
-            "Do not include any extra text or markdown—only return a valid Python dict literal."
+            "You are an agent that decides whether to call a tool or answer directly.\n"
+            "User’s question: \"" + user_input + "\"\n\n"
+            "• If you can answer directly, return exactly {\"final_answer\":\"<your answer>\"}.\n"
+            "• Otherwise, respond with exactly one of:\n"
+            "    {\"web_search_query\":\"<search terms>\"}\n"
+            "    {\"ocr_path\":\"<path to image>\"}\n"
+            "    {\"excel_path\":\"<path to xlsx>\", \"excel_sheet_name\":\"<sheet>\"}\n"
+            "    {\"audio_path\":\"<path to audio file>\"}\n"
+            "Do not include any extra characters or markdown—only the JSON literal."
         )
     )
     human_msg = HumanMessage(content=user_input)
 
-    # 3) Call the LLM with a brand‐new list [system_msg, human_msg]
+    # (2) Call the LLM
     llm_response = llm([system_msg, human_msg])
     llm_out = llm_response.content.strip()
 
-    # 4) Always append the LLM output as an AIMessage
+    # ── DEBUG: print raw LLM output ──
+    print("\n>>> plan_node got raw LLM output:")
+    print(llm_out)
+    print("<<< end raw output\n")
+
+    # (3) Append the LLM output to the message history
     ai_msg = AIMessage(content=llm_out)
     new_msgs = prior_msgs.copy() + [ai_msg]
+
+    # (4) Try parsing as JSON
     try:
-        parsed = eval(llm_out, {}, {})
+        parsed = json.loads(llm_out)
+        print(">>> plan_node parsed JSON:", parsed)
         if isinstance(parsed, dict):
             partial: AgentState = {"messages": new_msgs}
             allowed = {
@@ -78,31 +81,29 @@ def plan_node(state: AgentState) -> AgentState:
             for k, v in parsed.items():
                 if k in allowed:
                     partial[k] = v
+                    print(f">>> plan_node is setting {k!r} → {v!r}")
             return partial
-    except Exception:
-        pass
+    except json.JSONDecodeError as e:
+        print(">>> plan_node JSON parse error:", e)
+
+    # (5) Fallback
+    print(">>> plan_node falling back to final_answer alone\n")
+    return {"messages": new_msgs, "final_answer": "Sorry, I could not parse your intent."}
 
-    # 5) Fallback
-    return {
-        "messages": new_msgs,
-        "final_answer": "Sorry, I could not parse your intent."
-    }
 
 
 # ─── 3) Revised finalize_node ───
 def finalize_node(state: AgentState) -> AgentState:
-    # If plan_node already provided a final answer, skip LLM
     if state.get("final_answer") is not None:
+        print(">>> finalize_node: returning existing final_answer:", state["final_answer"])
         return {"final_answer": state["final_answer"]}
 
-    # Re-extract the last user question from messages
     question = ""
     for msg in reversed(state.get("messages", [])):
         if isinstance(msg, HumanMessage):
             question = msg.content
             break
 
-    # Build a combined context
     combined = f"USER_QUESTION: {question}\n"
     if sr := state.get("web_search_result"):
         combined += f"WEB_SEARCH_RESULT: {sr}\n"
@@ -110,11 +111,17 @@ def finalize_node(state: AgentState) -> AgentState:
         combined += f"OCR_RESULT: {orc}\n"
     if exr := state.get("excel_result"):
         combined += f"EXCEL_RESULT: {exr}\n"
-    # Check for both possible transcript keys
     audio_transcript = state.get("audio_transcript") or state.get("transcript")
     if audio_transcript:
         combined += f"AUDIO_TRANSCRIPT: {audio_transcript}\n"
-    combined += "Based on the above, provide ONLY the final answer. Do not include any explanation or extra text. the final answer should be in the same language as the user's question. should be as few words as possible. if its multiple words, put them comma sepreated. if its a number, then return the number ONLY as a string."
+
+    combined += (
+        "Based on the above, provide ONLY the final answer. "
+        "…(rest of your instructions)…"
+    )
+
+    # DEBUG: show exactly what we're sending to GPT-4 for final answer
+    print("\n>>> finalize_node prompt to LLM:\n" + combined + "\n<<< end prompt >>>\n")
 
     llm_response = llm([SystemMessage(content=combined)])
     return {"final_answer": llm_response.content.strip()}
@@ -153,17 +160,21 @@ graph.add_edge(START, "plan")
 
 # 5.c) plan → conditional: if any tool key was set, go to "tools"; otherwise "finalize"
 def route_plan(plan_out: AgentState) -> str:
+    # print what keys are present in plan_out
+    print(f">> route_plan sees plan_out keys: {list(plan_out.keys())}")
+
     if (
         plan_out.get("web_search_query")
         or plan_out.get("ocr_path")
         or plan_out.get("excel_path")
-        or plan_out.get("audio_path")     # newly added
+        or plan_out.get("audio_path")
     ):
-        print("Going to tools")
+        print(">> route_plan ➡️ tools")
         return "tools"
-    print("Going to finalize")
+    print(">> route_plan ➡️ finalize")
     return "finalize"
 
+
 graph.add_conditional_edges(
     "plan",
     route_plan,