Update app.py

app.py

@@ -160,23 +160,78 @@ def inspect_node(state: AgentState) -> AgentState:
     After running a tool, show GPT:
       - ORIGINAL user question
       - Any tool results (web_search_result, ocr_result, excel_result, transcript, wiki_result)
-      - The INTERIM_ANSWER (what plan_node initially provided under 'final_answer')
-    Then ask GPT to either:
+      - The INTERIM_ANSWER (always present if plan_node ran correctly)
+
+    If tool_counter ≥ 5, use LLM once more (with full context) to craft a final answer.
+    Otherwise, ask GPT to either:
       • Return {"final_answer":"<final>"} if done, OR
       • Return exactly one tool key to run next (wiki_query / ocr_path / excel_path & excel_sheet_name / audio_path).
     """
     global tool_counter
 
-    # 0) If we've already called tools too many times, force a final answer:
+    # 0) MAX‐TOOLS GUARD: if we've already run 5 tools, ask LLM to finalize
     if tool_counter >= 5:
-        return {
-            "messages": state["messages"],
-            "final_answer": state.get("final_answer", "ERROR: no interim_answer to finalize."),
-        }
+        # Build a fresh prompt that includes:
+        #  - The user’s original question
+        #  - Any tool results collected so far
+        #  - The interim answer
+        # Then ask for a polished final_answer.
+        messages_for_llm = []
+
+        # (1) Put back the user question
+        question = ""
+        for msg in reversed(state.get("messages", [])):
+            if isinstance(msg, HumanMessage):
+                question = msg.content
+                break
+        messages_for_llm.append(SystemMessage(content=f"USER_QUESTION: {question}"))
+
+        # (2) Add any tool results
+        if sr := state.get("web_search_result"):
+            messages_for_llm.append(SystemMessage(content=f"WEB_SEARCH_RESULT: {sr}"))
+        if orc := state.get("ocr_result"):
+            messages_for_llm.append(SystemMessage(content=f"OCR_RESULT: {orc}"))
+        if exr := state.get("excel_result"):
+            messages_for_llm.append(SystemMessage(content=f"EXCEL_RESULT: {exr}"))
+        if tr := state.get("transcript"):
+            messages_for_llm.append(SystemMessage(content=f"AUDIO_TRANSCRIPT: {tr}"))
+        if wr := state.get("wiki_result"):
+            messages_for_llm.append(SystemMessage(content=f"WIKIPEDIA_RESULT: {wr}"))
+
+        # (3) Always show the interim answer
+        interim = state.get("interim_answer", "")
+        messages_for_llm.append(SystemMessage(content=f"INTERIM_ANSWER: {interim}"))
+
+        # (4) Prompt GPT to produce a polished final answer
+        prompt = (
+            "You have the user’s question, any relevant tool results above,"
+            " and a draft answer (INTERIM_ANSWER).\n"
+            "Using all of that, write a concise, polished final answer. "
+            "Return exactly:\n"
+            "  {\"final_answer\":\"<your final answer>\"}\n"
+            "and nothing else.\n"
+        )
+        messages_for_llm.append(SystemMessage(content=prompt))
+
+        llm_response = llm(messages_for_llm)
+        raw = llm_response.content.strip()
+        new_msgs = state["messages"] + [AIMessage(content=raw)]
 
+        # Parse the LLM’s JSON
+        try:
+            parsed = json.loads(raw)
+            if isinstance(parsed, dict) and "final_answer" in parsed:
+                return {"messages": new_msgs, "final_answer": parsed["final_answer"]}
+        except json.JSONDecodeError:
+            pass
+
+        # If parsing fails, at least return the interim as fallback
+        return {"messages": new_msgs, "final_answer": interim}
+
+    # ——————————— If tool_counter < 5, proceed as before ———————————
     messages_for_llm = []
 
-    # 1) Re‐insert original user question
+    # (1) Re‐insert original user question
     question = ""
     for msg in reversed(state.get("messages", [])):
         if isinstance(msg, HumanMessage):
@@ -184,7 +239,7 @@ def inspect_node(state: AgentState) -> AgentState:
             break
     messages_for_llm.append(SystemMessage(content=f"USER_QUESTION: {question}"))
 
-    # 2) Add any tool results
+    # (2) Add any tool results
     if sr := state.get("web_search_result"):
         messages_for_llm.append(SystemMessage(content=f"WEB_SEARCH_RESULT: {sr}"))
     if orc := state.get("ocr_result"):
@@ -196,14 +251,14 @@ def inspect_node(state: AgentState) -> AgentState:
     if wr := state.get("wiki_result"):
         messages_for_llm.append(SystemMessage(content=f"WIKIPEDIA_RESULT: {wr}"))
 
-    # 3) Add the interim answer under INTERIM_ANSWER
-    if ia := state.get("final_answer"):
-        messages_for_llm.append(SystemMessage(content=f"INTERIM_ANSWER: {ia}"))
+    # (3) Always show the interim answer
+    interim = state.get("interim_answer", "")
+    messages_for_llm.append(SystemMessage(content=f"INTERIM_ANSWER: {interim}"))
 
-    # 4) Prompt GPT to decide final or another tool
+    # (4) Prompt GPT to decide final or another tool
     prompt = (
         "You have a current draft answer (INTERIM_ANSWER) and possibly some tool results above.\n"
-        "If you are confident it's correct, return exactly:\n"
+        "If you are confident it’s correct, return exactly:\n"
         "  {\"final_answer\":\"<your final answer>\"}\n"
         "and nothing else.\n"
         "Otherwise, return exactly one of these JSON literals to fetch another tool:\n"
@@ -216,35 +271,30 @@ def inspect_node(state: AgentState) -> AgentState:
     messages_for_llm.append(SystemMessage(content=prompt))
     llm_response = llm(messages_for_llm)
     raw = llm_response.content.strip()
-
     new_msgs = state["messages"] + [AIMessage(content=raw)]
 
-    # Try to parse the LLM's JSON
+    # Try to parse the LLM’s JSON
     try:
         parsed = json.loads(raw)
         if isinstance(parsed, dict):
-            # If GPT gave a final_answer, we finish here
+            # (a) If GPT gave a final_answer, return immediately
             if "final_answer" in parsed:
                 return {"messages": new_msgs, "final_answer": parsed["final_answer"]}
 
-            # If GPT requested exactly one valid tool, return only that key
+            # (b) If GPT requested exactly one valid tool, return only that key
             valid_keys = {"wiki_query", "ocr_path", "excel_path", "excel_sheet_name", "audio_path"}
             requested_keys = set(parsed.keys()) & valid_keys
             if len(requested_keys) == 1:
                 clean: AgentState = {"messages": new_msgs}
-                # Carry forward the global tool_counter implicitly (no need to store in state)
                 for k in requested_keys:
                     clean[k] = parsed[k]
                 return clean
     except json.JSONDecodeError:
         pass
 
-    # Fallback: if GPT didn't give a valid tool key or final_answer, finalize with existing interim
-    if ia := state.get("final_answer"):
-        return {"messages": new_msgs, "final_answer": ia}
-
-    # If there is no interim either, we cannot proceed
-    return {"messages": new_msgs, "final_answer": "ERROR: could not parse inspect decision."}
+    # (c) Fallback: if GPT never returned a valid tool key or a final_answer,
+    # just finalize with the existing interim_answer
+    return {"messages": new_msgs, "final_answer": interim}
 
 
 # ─── 6) finalize_node ───