version2

app.py

@@ -22,17 +22,20 @@ DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
 
 from tools import ocr_image_tool, parse_excel_tool, web_search_tool, run_tools, audio_transcriber_tool, wikipedia_search_tool
 
-llm = ChatOpenAI(model_name="gpt-4.1")
-
-# agent = create_react_agent(model=llm, tools=tool_node)
+llm = ChatOpenAI(model_name="gpt-4o-mini")
 
+# ─── 1) plan_node ───
 def plan_node(state: AgentState) -> AgentState:
     """
-    This plan_node will ask GPT to:
-      1) First write a concise *direct* answer.
-      2) Then decide whether it’s confident enough to stop (return {"final_answer": ...})
-         or if it needs to verify via one tool (return exactly one of {"wiki_query":...}, 
-         {"web_search_query":...}, {"ocr_path":...}, {"excel_path":...,"excel_sheet_name":...}, or {"audio_path":...}).
+    Step 1: Ask GPT to draft a concise direct answer (INTERIM_ANSWER),
+            then decide if it’s confident enough to stop or if it needs one tool.
+    If confident: return {"final_answer": "<answer>"}
+    Otherwise:   return exactly one of 
+                 {"wiki_query": "..."},
+                 {"web_search_query": "..."},
+                 {"ocr_path": "..."},
+                 {"excel_path": "...", "excel_sheet_name": "..."},
+                 {"audio_path": "..."}
     """
     prior_msgs = state.get("messages", [])
     user_input = ""
@@ -41,38 +44,32 @@ def plan_node(state: AgentState) -> AgentState:
             user_input = msg.content
             break
 
-    # (1) Build a fresh SystemMessage that tells the LLM exactly how to self‐evaluate
     system_msg = SystemMessage(
         content=(
-            "You are an agent that must do two things in a single JSON output:\n\n"
-            "  1) Produce a concise, direct answer to the user’s question (no explanation, just the answer).  \n"
-            "  2) Judge whether that answer is reliable.  \n"
-            "     • If you are fully confident and do NOT need any external verification, return exactly:\n"
+            "You are an agent that must do two things in one JSON output:\n\n"
+            "  1) Provide a concise, direct answer to the user’s question (no explanation).\n"
+            "  2) Judge whether that answer is reliable:\n"
+            "     • If you are fully confident, return exactly:\n"
             "         {\"final_answer\":\"<your concise answer>\"}\n"
             "       and nothing else.\n"
-            "     • If you think you need to verify or look something up first, return exactly one of the following (and nothing else):\n"
-            "         {\"wiki_query\":\"<search terms for Wikipedia>\"}\n"
+            "     • Otherwise, return exactly one of:\n"
+            "         {\"wiki_query\":\"<Wikipedia search>\"}\n"
             "         {\"web_search_query\":\"<search terms>\"}\n"
-            "         {\"ocr_path\":\"<local image path or task_id>\"}\n"
-            "         {\"excel_path\":\"<local .xlsx path>\", \"excel_sheet_name\":\"<sheet name>\"}\n"
-            "         {\"audio_path\":\"<local audio path or task_id>\"}\n\n"
-            "       You must pick exactly one key—either final_answer or exactly one tool key.\n"
-            "       Do NOT wrap it in any markdown or extra text.  Only output a single JSON object.\n"
-            "\n"
+            "         {\"ocr_path\":\"<image path or task_id>\"}\n"
+            "         {\"excel_path\":\"<xlsx path>\", \"excel_sheet_name\":\"<sheet name>\"}\n"
+            "         {\"audio_path\":\"<audio path or task_id>\"}\n"
+            "       and nothing else.\n"
+            "Do NOT wrap in markdown—output only a single JSON object.\n"
             f"User’s question: \"{user_input}\"\n"
         )
     )
     human_msg = HumanMessage(content=user_input)
-
-    # (2) Call the LLM with this single system/human pair
     llm_response = llm([system_msg, human_msg])
     llm_out = llm_response.content.strip()
 
-    # (3) Append the LLM output into the message history
     ai_msg = AIMessage(content=llm_out)
     new_msgs = prior_msgs.copy() + [ai_msg]
 
-    # (4) Attempt to parse that JSON
     try:
         parsed = json.loads(llm_out)
         if isinstance(parsed, dict):
@@ -93,208 +90,221 @@ def plan_node(state: AgentState) -> AgentState:
     except json.JSONDecodeError:
         pass
 
-    # (5) If parsing failed, fall back to a safe “sorry” answer
     return {
         "messages": new_msgs,
         "final_answer": "Sorry, I could not parse your intent."
     }
 
 
+# ─── 2) store_prev_state ───
+def store_prev_state(state: AgentState) -> AgentState:
+    return {**state, "prev_state": state.copy()}
 
-# ─── 3) Revised finalize_node ───
-def finalize_node(state: AgentState) -> AgentState:
-    if state.get("final_answer") is not None:
-        return {"final_answer": state["final_answer"]}
 
-    # Re‐extract the last user question
+# ─── 3) tools_node ───
+def tool_node(state: AgentState) -> AgentState:
+    """
+    Dispatch exactly one tool based on which key was set:
+      - wiki_query → wikipedia_search_tool
+      - web_search_query → web_search_tool
+      - ocr_path → ocr_image_tool
+      - excel_path → parse_excel_tool
+      - audio_path → audio_transcriber_tool
+    """
+    if state.get("wiki_query"):
+        return wikipedia_search_tool(state)
+    if state.get("web_search_query"):
+        return web_search_tool(state)
+    if state.get("ocr_path"):
+        return ocr_image_tool(state)
+    if state.get("excel_path"):
+        return parse_excel_tool(state)
+    if state.get("audio_path"):
+        return audio_transcriber_tool(state)
+    return {}
+
+
+# ─── 4) merge_tool_output ───
+def merge_tool_output(state: AgentState) -> AgentState:
+    """
+    Combine previous state and tool output into one:
+    """
+    prev = state.get("prev_state", {})
+    merged = {**prev, **state}
+    merged.pop("prev_state", None)
+    return merged
+
+
+# ─── 5) inspect_node ───
+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:
+      • Return {"final_answer": "<final>"} if done, OR
+      • Return exactly one tool key to run next (wiki_query / web_search_query / ocr_path / excel_path & excel_sheet_name / audio_path).
+    """
+    messages_for_llm = []
+
+    # 1) Re‐insert original 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}"))
 
-    # Build one monolithic context
-    combined = f"USER_QUESTION: {question}\n"
+    # 2) Add any tool results
     if sr := state.get("web_search_result"):
-        combined += f"WEB_SEARCH_RESULT: {sr}\n"
+        messages_for_llm.append(SystemMessage(content=f"WEB_SEARCH_RESULT: {sr}"))
     if orc := state.get("ocr_result"):
-        combined += f"OCR_RESULT: {orc}\n"
+        messages_for_llm.append(SystemMessage(content=f"OCR_RESULT: {orc}"))
     if exr := state.get("excel_result"):
-        combined += f"EXCEL_RESULT: {exr}\n"
-    # Note: your code already stores the audio transcription under "transcript"
+        messages_for_llm.append(SystemMessage(content=f"EXCEL_RESULT: {exr}"))
     if tr := state.get("transcript"):
-        combined += f"AUDIO_TRANSCRIPT: {tr}\n"
+        messages_for_llm.append(SystemMessage(content=f"AUDIO_TRANSCRIPT: {tr}"))
     if wr := state.get("wiki_result"):
-        combined += f"WIKIPEDIA_RESULT: {wr}\n"
-
-    # Here we demand a JSON response with a single key "final_answer"
-    combined += (
-        "Based on the above, respond with exactly one JSON object, and nothing else. "
-        "The JSON object must have exactly one key: \"final_answer\". "
-        "For example:\n"
-        "{\"final_answer\":\"42\"}\n"
-        "Do NOT include any explanation, markdown, or any extra whitespace outside the JSON object. "
-        "If the answer is multiple words, put them in a comma-separated string, e.g. \"red,green,blue\". "
-        "If the answer is a number, it must be digits only—e.g. \"725.00\".\n"
-        "If the answer is a list of items, put them in a comma-separated string, e.g. \"item1,item2,item3\". "
-        "If the user prompt asks you to do something, then do it "
+        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}"))
+
+    # 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"
+        "  {\"final_answer\":\"<your final answer>\"}\n"
+        "and nothing else.\n"
+        "Otherwise, return exactly one of these JSON literals to fetch another tool:\n"
+        "  {\"wiki_query\":\"<query for Wikipedia>\"}\n"
+        "  {\"web_search_query\":\"<search terms>\"}\n"
+        "  {\"ocr_path\":\"<image path or task_id>\"}\n"
+        "  {\"excel_path\":\"<xls path>\", \"excel_sheet_name\":\"<sheet name>\"}\n"
+        "  {\"audio_path\":\"<audio path or task_id>\"}\n"
+        "Do NOT wrap in markdown—return only the JSON object.\n"
     )
-
-    # Debug print
-    # print("\n>>> finalize_node JSON‐strict prompt:\n" + combined + "\n<<< end prompt >>>\n")
-
-    llm_response = llm.invoke([SystemMessage(content=combined)])
+    messages_for_llm.append(SystemMessage(content=prompt))
+    llm_response = llm(messages_for_llm)
     raw = llm_response.content.strip()
-    # print(">>> finalize_node got raw response:", raw)
 
+    new_msgs = state["messages"] + [AIMessage(content=raw)]
     try:
         parsed = json.loads(raw)
-        return {"final_answer": parsed["final_answer"]}
-    except Exception as e:
-        # If the LLM did not return valid JSON, store the error so you can see it
-        # print(">>> finalize_node JSON parse error:", e, "raw was:", raw)
-        return {"final_answer": f"ERROR: invalid JSON from finalize_node: {raw}"}
-
-# ─── 4) Wrap tools in a ToolNode ───
-def tool_node(state: AgentState) -> AgentState:
-    """
-    Inspect exactly which tool‐key was set in `state` and call that function.
-    Returns only the partial state (with the tool's outputs) so that merge_tool_output can combine it.
-    """
-    # We expect exactly one of these keys to be non‐empty:
-    #   "web_search_query", "ocr_path", "excel_path"/"excel_sheet_name", "audio_path"
-    # Whichever is present, call the corresponding tool and return its result.
-    
-    if state.get("wiki_query"):
-        out = wikipedia_search_tool(state)
-        return out
-    
-    if state.get("web_search_query"):
-        # print(f">>> tools_node dispatching web_search_tool with query: {state['web_search_query']!r}")
-        out = web_search_tool(state)
-        return out
-
-    if state.get("ocr_path"):
-        # print(f">>> tools_node dispatching ocr_image_tool with path: {state['ocr_path']!r}")
-        out = ocr_image_tool(state)
-        return out
-
-    if state.get("excel_path"):
-        # We assume plan_node always sets both excel_path and excel_sheet_name together
-        # print(f">>> tools_node dispatching parse_excel_tool with path: {state['excel_path']!r}, sheet: {state.get('excel_sheet_name')!r}")
-        out = parse_excel_tool(state)
-        return out
-
-    if state.get("audio_path"):
-        # print(f">>> tools_node dispatching audio_transcriber_tool with path: {state['audio_path']!r}")
-        out = audio_transcriber_tool(state)
-        return out
-
-
-
-    # If we somehow reach here, no recognized tool key was set:
-    # print(">>> tools_node: no valid tool key found in state!")
-    return {}
+        if isinstance(parsed, dict):
+            partial: AgentState = {"messages": new_msgs}
+            allowed = {
+                "final_answer",
+                "wiki_query",
+                "web_search_query",
+                "ocr_path",
+                "excel_path",
+                "excel_sheet_name",
+                "audio_path"
+            }
+            for k, v in parsed.items():
+                if k in allowed:
+                    partial[k] = v
+            return partial
+    except json.JSONDecodeError:
+        pass
 
+    return {
+        "messages": new_msgs,
+        "final_answer": "ERROR: could not parse inspect decision."
+    }
 
-# Add a node to store the previous state
 
-def store_prev_state(state: AgentState) -> AgentState:
-    return {**state, "prev_state": state.copy()}
+# ─── 6) finalize_node ───
+def finalize_node(state: AgentState) -> AgentState:
+    """
+    If state already has "final_answer", return it. Otherwise, gather all tool outputs
+    and ask GPT for a final answer. But in our cyclic design, finalize_node is only called
+    after plan_node or inspect_node returned "final_answer".
+    """
+    if fa := state.get("final_answer"):
+        return {"final_answer": fa}
+    # (In practice, we never reach here because we always pick finalize only when "final_answer" exists.)
+    return {"final_answer": "ERROR: finalize called without a final_answer."}
 
-def merge_tool_output(state: AgentState) -> AgentState:
-    prev_state = state.get("prev_state", {})
-    merged = {**prev_state, **state}
-    merged.pop("prev_state", None)
-    return merged
 
-# ─── 5) Build the graph ───
+# ─── 7) Build the graph and wire edges ───
 graph = StateGraph(AgentState)
 
-# 5.a) Register nodes
+# Register nodes
 graph.add_node("plan", plan_node)
 graph.add_node("store_prev_state", store_prev_state)
 graph.add_node("tools", tool_node)
 graph.add_node("merge_tool_output", merge_tool_output)
+graph.add_node("inspect", inspect_node)
 graph.add_node("finalize", finalize_node)
 
-# 5.b) Wire START → plan
+# START → plan
 graph.add_edge(START, "plan")
 
-# 5.c) plan → conditional: if any tool key was set, go to "tools"; otherwise "finalize"
+# plan → either finalize (if plan set final_answer) or store_prev_state (if plan wants a tool)
 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")
-        or plan_out.get("wiki_query")
-    ):
-        # print(">> route_plan ➡️ tools")
-        return "tools"
-    # print(">> route_plan ➡️ finalize")
-    return "finalize"
-
+    if plan_out.get("final_answer") is not None:
+        return "finalize"
+    return "store_prev_state"
 
 graph.add_conditional_edges(
     "plan",
     route_plan,
-    {"tools": "store_prev_state", "finalize": "finalize"}
+    {"store_prev_state": "store_prev_state", "finalize": "finalize"}
 )
 
-# 5.d) store_prev_state → tools
+# store_prev_state → tools
 graph.add_edge("store_prev_state", "tools")
 
-# 5.e) tools → merge_tool_output
+# tools → merge_tool_output
 graph.add_edge("tools", "merge_tool_output")
 
-# 5.f) merge_tool_output → finalize
-graph.add_edge("merge_tool_output", "finalize")
+# merge_tool_output → inspect
+graph.add_edge("merge_tool_output", "inspect")
+
+# inspect → either finalize (if inspect set final_answer) or store_prev_state (if inspect wants another tool)
+def route_inspect(inspect_out: AgentState) -> str:
+    if inspect_out.get("final_answer") is not None:
+        return "finalize"
+    return "store_prev_state"
+
+graph.add_conditional_edges(
+    "inspect",
+    route_inspect,
+    {"store_prev_state": "store_prev_state", "finalize": "finalize"}
+)
 
-# 5.g) finalize → END
+# finalize → END
 graph.add_edge("finalize", END)
 
 compiled_graph = graph.compile()
 
 
-# ─── 6) respond_to_input ───
-def respond_to_input(user_input: str, task_id) -> str:
+# ─── 8) respond_to_input ───
+def respond_to_input(user_input: str) -> str:
     """
-    Seed state['messages'] with a SystemMessage (tools description) + HumanMessage(user_input).
-    Then invoke the graph; return the final_answer from the resulting state.
+    Seed state['messages'] with a SystemMessage + HumanMessage(user_input),
+    then invoke the cyclic graph. Return the final_answer from the resulting state.
     """
     system_msg = SystemMessage(
-    content=(
-        "You are an agent that must choose exactly one of the following actions:\n"
-        "  1) If the user's question can be answered directly by consulting Wikipedia, return exactly:\n"
-        "       {\"wiki_query\":\"<search terms for Wikipedia>\"}\n"
-        "     and nothing else. Use Wikipedia before any other tool.\n"
-        "  2) Only if Wikipedia cannot directly answer, perform a web search and return:\n"
-        "       {\"web_search_query\":\"<search terms>\"}\n"
-        "     and nothing else.\n"
-        "  3) If the user's question requires extracting text from an image, return:\n"
-        "       {\"ocr_path\":\"<local image path>\"}\n"
-        "     and nothing else.\n"
-        "  4) If the user's question requires reading a spreadsheet, return:\n"
-        "       {\"excel_path\":\"<local .xlsx path>\", \"excel_sheet_name\":\"<sheet name>\"}\n"
-        "     and nothing else.\n"
-        "  5) If the user needs an audio transcription, return:\n"
-        "       {\"audio_path\":\"<local audio file path>\"}\n"
-        "     and nothing else.\n"
-        "  6) If you already know the answer without using any tool, return exactly:\n"
-        "       {\"final_answer\":\"<your concise answer>\"}\n"
-        "     and nothing else.\n"
-        "If the user's prompt explicitly tells you to perform a specific action (for example, “translate this sentence”), then do it directly and return your result as {\"final_answer\":\"<your answer>\"} or the appropriate tool key if needed.  \n"
-        "Do NOT include any additional keys, explanation, or markdown—only one JSON object with exactly one key."
+        content=(
+            "You are an agent orchestrator. Decide whether to use a tool or answer directly.\n"
+            "Tools available:\n"
+            "  • Wikipedia: set {\"wiki_query\":\"<search terms>\"}\n"
+            "  • Web search: set {\"web_search_query\":\"<search terms>\"}\n"
+            "  • OCR: set {\"ocr_path\":\"<image path or task_id>\"}\n"
+            "  • Excel: set {\"excel_path\":\"<xlsx path>\", \"excel_sheet_name\":\"<sheet>\"}\n"
+            "  • Audio transcription: set {\"audio_path\":\"<audio path or task_id>\"}\n"
+            "If you can answer immediately, set {\"final_answer\":\"<answer>\"}. "
+            "Respond with only one JSON object and no extra formatting."
+        )
     )
-)
-
     human_msg = HumanMessage(content=user_input)
 
-    initial_state: AgentState = {"messages": [system_msg, human_msg], "task_id": task_id}
+    initial_state: AgentState = {"messages": [system_msg, human_msg]}
     final_state = compiled_graph.invoke(initial_state)
     return final_state.get("final_answer", "Error: No final answer generated.")
 

old_app_copy.py

@@ -0,0 +1,504 @@
+import os
+import gradio as gr
+import requests
+import inspect
+import pandas as pd
+from langgraph.prebuilt import ToolNode
+
+
+# from typing import Any, Dict
+# from typing import TypedDict, Annotated
+
+from langchain_openai import ChatOpenAI
+from langgraph.graph import StateGraph, START, END
+from langgraph.graph.message import add_messages
+from langchain.schema import HumanMessage, SystemMessage, AIMessage
+# Create a ToolNode that knows about your web_search function
+import json
+from state import AgentState
+
+# --- Constants ---
+DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
+
+from tools import ocr_image_tool, parse_excel_tool, web_search_tool, run_tools, audio_transcriber_tool, wikipedia_search_tool
+
+llm = ChatOpenAI(model_name="gpt-4o-mini")
+
+# agent = create_react_agent(model=llm, tools=tool_node)
+
+def plan_node(state: AgentState) -> AgentState:
+    """
+    This plan_node will ask GPT to:
+      1) First write a concise *direct* answer.
+      2) Then decide whether it’s confident enough to stop (return {"final_answer": ...})
+         or if it needs to verify via one tool (return exactly one of {"wiki_query":...}, 
+         {"web_search_query":...}, {"ocr_path":...}, {"excel_path":...,"excel_sheet_name":...}, or {"audio_path":...}).
+    """
+    prior_msgs = state.get("messages", [])
+    user_input = ""
+    for msg in reversed(prior_msgs):
+        if isinstance(msg, HumanMessage):
+            user_input = msg.content
+            break
+
+    # (1) Build a fresh SystemMessage that tells the LLM exactly how to self‐evaluate
+    system_msg = SystemMessage(
+        content=(
+            "You are an agent that must do two things in a single JSON output:\n\n"
+            "  1) Produce a concise, direct answer to the user’s question (no explanation, just the answer).  \n"
+            "  2) Judge whether that answer is reliable.  \n"
+            "     • If you are fully confident and do NOT need any external verification, return exactly:\n"
+            "         {\"final_answer\":\"<your concise answer>\"}\n"
+            "       and nothing else.\n"
+            "     • If you think you need to verify or look something up first, return exactly one of the following (and nothing else):\n"
+            "         {\"wiki_query\":\"<search terms for Wikipedia>\"}\n"
+            "         {\"web_search_query\":\"<search terms>\"}\n"
+            "         {\"ocr_path\":\"<local image path or task_id>\"}\n"
+            "         {\"excel_path\":\"<local .xlsx path>\", \"excel_sheet_name\":\"<sheet name>\"}\n"
+            "         {\"audio_path\":\"<local audio path or task_id>\"}\n\n"
+            "       You must pick exactly one key—either final_answer or exactly one tool key.\n"
+            "       Do NOT wrap it in any markdown or extra text.  Only output a single JSON object.\n"
+            "\n"
+            f"User’s question: \"{user_input}\"\n"
+        )
+    )
+    human_msg = HumanMessage(content=user_input)
+
+    # (2) Call the LLM with this single system/human pair
+    llm_response = llm([system_msg, human_msg])
+    llm_out = llm_response.content.strip()
+
+    # (3) Append the LLM output into the message history
+    ai_msg = AIMessage(content=llm_out)
+    new_msgs = prior_msgs.copy() + [ai_msg]
+
+    # (4) Attempt to parse that JSON
+    try:
+        parsed = json.loads(llm_out)
+        if isinstance(parsed, dict):
+            partial: AgentState = {"messages": new_msgs}
+            allowed_keys = {
+                "final_answer",
+                "wiki_query",
+                "web_search_query",
+                "ocr_path",
+                "excel_path",
+                "excel_sheet_name",
+                "audio_path"
+            }
+            for k, v in parsed.items():
+                if k in allowed_keys:
+                    partial[k] = v
+            return partial
+    except json.JSONDecodeError:
+        pass
+
+    # (5) If parsing failed, fall back to a safe “sorry” answer
+    return {
+        "messages": new_msgs,
+        "final_answer": "Sorry, I could not parse your intent."
+    }
+
+
+
+# ─── 3) Revised finalize_node ───
+def finalize_node(state: AgentState) -> AgentState:
+    if state.get("final_answer") is not None:
+        return {"final_answer": state["final_answer"]}
+
+    # Re‐extract the last user question
+    question = ""
+    for msg in reversed(state.get("messages", [])):
+        if isinstance(msg, HumanMessage):
+            question = msg.content
+            break
+
+    # Build one monolithic 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"
+    # Note: your code already stores the audio transcription under "transcript"
+    if tr := state.get("transcript"):
+        combined += f"AUDIO_TRANSCRIPT: {tr}\n"
+    if wr := state.get("wiki_result"):
+        combined += f"WIKIPEDIA_RESULT: {wr}\n"
+
+    # Here we demand a JSON response with a single key "final_answer"
+    combined += (
+        "Based on the above, respond with exactly one JSON object, and nothing else. "
+        "The JSON object must have exactly one key: \"final_answer\". "
+        "For example:\n"
+        "{\"final_answer\":\"42\"}\n"
+        "Do NOT include any explanation, markdown, or any extra whitespace outside the JSON object. "
+        "If the answer is multiple words, put them in a comma-separated string, e.g. \"red,green,blue\". "
+        "If the answer is a number, it must be digits only—e.g. \"725.00\".\n"
+        "If the answer is a list of items, put them in a comma-separated string, e.g. \"item1,item2,item3\". "
+        "If the user prompt asks you to do something, then do it "
+    )
+
+    # Debug print
+    # print("\n>>> finalize_node JSON‐strict prompt:\n" + combined + "\n<<< end prompt >>>\n")
+
+    llm_response = llm.invoke([SystemMessage(content=combined)])
+    raw = llm_response.content.strip()
+    # print(">>> finalize_node got raw response:", raw)
+
+    try:
+        parsed = json.loads(raw)
+        return {"final_answer": parsed["final_answer"]}
+    except Exception as e:
+        # If the LLM did not return valid JSON, store the error so you can see it
+        # print(">>> finalize_node JSON parse error:", e, "raw was:", raw)
+        return {"final_answer": f"ERROR: invalid JSON from finalize_node: {raw}"}
+
+# ─── 4) Wrap tools in a ToolNode ───
+def tool_node(state: AgentState) -> AgentState:
+    """
+    Inspect exactly which tool‐key was set in `state` and call that function.
+    Returns only the partial state (with the tool's outputs) so that merge_tool_output can combine it.
+    """
+    # We expect exactly one of these keys to be non‐empty:
+    #   "web_search_query", "ocr_path", "excel_path"/"excel_sheet_name", "audio_path"
+    # Whichever is present, call the corresponding tool and return its result.
+    
+    if state.get("wiki_query"):
+        out = wikipedia_search_tool(state)
+        return out
+    
+    if state.get("web_search_query"):
+        # print(f">>> tools_node dispatching web_search_tool with query: {state['web_search_query']!r}")
+        out = web_search_tool(state)
+        return out
+
+    if state.get("ocr_path"):
+        # print(f">>> tools_node dispatching ocr_image_tool with path: {state['ocr_path']!r}")
+        out = ocr_image_tool(state)
+        return out
+
+    if state.get("excel_path"):
+        # We assume plan_node always sets both excel_path and excel_sheet_name together
+        # print(f">>> tools_node dispatching parse_excel_tool with path: {state['excel_path']!r}, sheet: {state.get('excel_sheet_name')!r}")
+        out = parse_excel_tool(state)
+        return out
+
+    if state.get("audio_path"):
+        # print(f">>> tools_node dispatching audio_transcriber_tool with path: {state['audio_path']!r}")
+        out = audio_transcriber_tool(state)
+        return out
+
+
+
+    # If we somehow reach here, no recognized tool key was set:
+    # print(">>> tools_node: no valid tool key found in state!")
+    return {}
+
+
+# Add a node to store the previous state
+
+def store_prev_state(state: AgentState) -> AgentState:
+    return {**state, "prev_state": state.copy()}
+
+def merge_tool_output(state: AgentState) -> AgentState:
+    prev_state = state.get("prev_state", {})
+    merged = {**prev_state, **state}
+    merged.pop("prev_state", None)
+    return merged
+
+# ─── 5) Build the graph ───
+graph = StateGraph(AgentState)
+
+# 5.a) Register nodes
+graph.add_node("plan", plan_node)
+graph.add_node("store_prev_state", store_prev_state)
+graph.add_node("tools", tool_node)
+graph.add_node("merge_tool_output", merge_tool_output)
+graph.add_node("finalize", finalize_node)
+
+# 5.b) Wire START → plan
+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")
+        or plan_out.get("wiki_query")
+    ):
+        # print(">> route_plan ➡️ tools")
+        return "tools"
+    # print(">> route_plan ➡️ finalize")
+    return "finalize"
+
+
+graph.add_conditional_edges(
+    "plan",
+    route_plan,
+    {"tools": "store_prev_state", "finalize": "finalize"}
+)
+
+# 5.d) store_prev_state → tools
+graph.add_edge("store_prev_state", "tools")
+
+# 5.e) tools → merge_tool_output
+graph.add_edge("tools", "merge_tool_output")
+
+# 5.f) merge_tool_output → finalize
+graph.add_edge("merge_tool_output", "finalize")
+
+# 5.g) finalize → END
+graph.add_edge("finalize", END)
+
+compiled_graph = graph.compile()
+
+
+# ���── 6) respond_to_input ───
+def respond_to_input(user_input: str, task_id) -> str:
+    """
+    Seed state['messages'] with a SystemMessage (tools description) + HumanMessage(user_input).
+    Then invoke the graph; return the final_answer from the resulting state.
+    """
+    system_msg = SystemMessage(
+    content=(
+        "You are an agent that must choose exactly one of the following actions:\n"
+        "  1) If the user's question can be answered directly by consulting Wikipedia, return exactly:\n"
+        "       {\"wiki_query\":\"<search terms for Wikipedia>\"}\n"
+        "     and nothing else. Use Wikipedia before any other tool.\n"
+        "  2) Only if Wikipedia cannot directly answer, perform a web search and return:\n"
+        "       {\"web_search_query\":\"<search terms>\"}\n"
+        "     and nothing else.\n"
+        "  3) If the user's question requires extracting text from an image, return:\n"
+        "       {\"ocr_path\":\"<local image path>\"}\n"
+        "     and nothing else.\n"
+        "  4) If the user's question requires reading a spreadsheet, return:\n"
+        "       {\"excel_path\":\"<local .xlsx path>\", \"excel_sheet_name\":\"<sheet name>\"}\n"
+        "     and nothing else.\n"
+        "  5) If the user needs an audio transcription, return:\n"
+        "       {\"audio_path\":\"<local audio file path>\"}\n"
+        "     and nothing else.\n"
+        "  6) If you already know the answer without using any tool, return exactly:\n"
+        "       {\"final_answer\":\"<your concise answer>\"}\n"
+        "     and nothing else.\n"
+        "If the user's prompt explicitly tells you to perform a specific action (for example, “translate this sentence”), then do it directly and return your result as {\"final_answer\":\"<your answer>\"} or the appropriate tool key if needed.  \n"
+        "Do NOT include any additional keys, explanation, or markdown—only one JSON object with exactly one key."
+    )
+)
+
+    human_msg = HumanMessage(content=user_input)
+
+    initial_state: AgentState = {"messages": [system_msg, human_msg], "task_id": task_id}
+    final_state = compiled_graph.invoke(initial_state)
+    return final_state.get("final_answer", "Error: No final answer generated.")
+
+
+
+
+class BasicAgent:
+    def __init__(self):
+        print("BasicAgent initialized.")
+    def __call__(self, question: str, task_id) -> str:
+        # print(f"Agent received question (first 50 chars): {question[:50]}...")
+        # fixed_answer = "This is a default answer."
+        # print(f"Agent returning fixed answer: {fixed_answer}")
+        print()
+        print()
+        print()
+        print()
+        
+        
+        print(f"Agent received question: {question}")
+        print()
+        return respond_to_input(question, task_id)
+        # return fixed_answer
+
+
+
+
+
+
+def run_and_submit_all( profile: gr.OAuthProfile | None):
+    """
+    Fetches all questions, runs the BasicAgent on them, submits all answers,
+    and displays the results.
+    """
+    # --- Determine HF Space Runtime URL and Repo URL ---
+    space_id = os.getenv("SPACE_ID") # Get the SPACE_ID for sending link to the code
+
+    if profile:
+        username= f"{profile.username}"
+        print(f"User logged in: {username}")
+    else:
+        print("User not logged in.")
+        return "Please Login to Hugging Face with the button.", None
+
+    api_url = DEFAULT_API_URL
+    questions_url = f"{api_url}/questions"
+    submit_url = f"{api_url}/submit"
+
+    # 1. Instantiate Agent ( modify this part to create your agent)
+    try:
+        agent = BasicAgent()
+    except Exception as e:
+        print(f"Error instantiating agent: {e}")
+        return f"Error initializing agent: {e}", None
+    # In the case of an app running as a hugging Face space, this link points toward your codebase ( usefull for others so please keep it public)
+    agent_code = f"https://huggingface.co/spaces/{space_id}/tree/main"
+    print(agent_code)
+
+    # 2. Fetch Questions
+    print(f"Fetching questions from: {questions_url}")
+    try:
+        response = requests.get(questions_url, timeout=15)
+        response.raise_for_status()
+        questions_data = response.json()
+        if not questions_data:
+             print("Fetched questions list is empty.")
+             return "Fetched questions list is empty or invalid format.", None
+        print(f"Fetched {len(questions_data)} questions.")
+    except requests.exceptions.RequestException as e:
+        print(f"Error fetching questions: {e}")
+        return f"Error fetching questions: {e}", None
+    except requests.exceptions.JSONDecodeError as e:
+         print(f"Error decoding JSON response from questions endpoint: {e}")
+         print(f"Response text: {response.text[:500]}")
+         return f"Error decoding server response for questions: {e}", None
+    except Exception as e:
+        print(f"An unexpected error occurred fetching questions: {e}")
+        return f"An unexpected error occurred fetching questions: {e}", None
+
+    # 3. Run your Agent
+    
+    results_log = []
+    answers_payload = []
+    print(f"Running agent on {len(questions_data)} questions...")
+    for item in questions_data:
+        task_id = item.get("task_id")
+        question_text = item.get("question")
+        if not task_id or question_text is None:
+            print(f"Skipping item with missing task_id or question: {item}")
+            continue
+        try:
+            submitted_answer = agent(question_text, task_id)
+            answers_payload.append({"task_id": task_id, "submitted_answer": submitted_answer})
+            results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": submitted_answer})
+        except Exception as e:
+             print(f"Error running agent on task {task_id}: {e}")
+             results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": f"AGENT ERROR: {e}"})
+
+    if not answers_payload:
+        print("Agent did not produce any answers to submit.")
+        return "Agent did not produce any answers to submit.", pd.DataFrame(results_log)
+
+    # 4. Prepare Submission 
+    submission_data = {"username": username.strip(), "agent_code": agent_code, "answers": answers_payload}
+    status_update = f"Agent finished. Submitting {len(answers_payload)} answers for user '{username}'..."
+    print(status_update)
+
+    # 5. Submit
+    print(f"Submitting {len(answers_payload)} answers to: {submit_url}")
+    try:
+        response = requests.post(submit_url, json=submission_data, timeout=60)
+        response.raise_for_status()
+        result_data = response.json()
+        final_status = (
+            f"Submission Successful!\n"
+            f"User: {result_data.get('username')}\n"
+            f"Overall Score: {result_data.get('score', 'N/A')}% "
+            f"({result_data.get('correct_count', '?')}/{result_data.get('total_attempted', '?')} correct)\n"
+            f"Message: {result_data.get('message', 'No message received.')}"
+        )
+        print("Submission successful.")
+        results_df = pd.DataFrame(results_log)
+        return final_status, results_df
+    except requests.exceptions.HTTPError as e:
+        error_detail = f"Server responded with status {e.response.status_code}."
+        try:
+            error_json = e.response.json()
+            error_detail += f" Detail: {error_json.get('detail', e.response.text)}"
+        except requests.exceptions.JSONDecodeError:
+            error_detail += f" Response: {e.response.text[:500]}"
+        status_message = f"Submission Failed: {error_detail}"
+        print(status_message)
+        results_df = pd.DataFrame(results_log)
+        return status_message, results_df
+    except requests.exceptions.Timeout:
+        status_message = "Submission Failed: The request timed out."
+        print(status_message)
+        results_df = pd.DataFrame(results_log)
+        return status_message, results_df
+    except requests.exceptions.RequestException as e:
+        status_message = f"Submission Failed: Network error - {e}"
+        print(status_message)
+        results_df = pd.DataFrame(results_log)
+        return status_message, results_df
+    except Exception as e:
+        status_message = f"An unexpected error occurred during submission: {e}"
+        print(status_message)
+        results_df = pd.DataFrame(results_log)
+        return status_message, results_df
+
+
+# --- Build Gradio Interface using Blocks ---
+with gr.Blocks() as demo:
+    gr.Markdown("# Basic Agent Evaluation Runner")
+    gr.Markdown(
+        """
+        **Instructions:**
+
+        1.  Please clone this space, then modify the code to define your agent's logic, the tools, the necessary packages, etc ...
+        2.  Log in to your Hugging Face account using the button below. This uses your HF username for submission.
+        3.  Click 'Run Evaluation & Submit All Answers' to fetch questions, run your agent, submit answers, and see the score.
+
+        ---
+        **Disclaimers:**
+        Once clicking on the "submit button, it can take quite some time ( this is the time for the agent to go through all the questions).
+        This space provides a basic setup and is intentionally sub-optimal to encourage you to develop your own, more robust solution. For instance for the delay process of the submit button, a solution could be to cache the answers and submit in a seperate action or even to answer the questions in async.
+        """
+    )
+
+    gr.LoginButton()
+
+    run_button = gr.Button("Run Evaluation & Submit All Answers")
+
+    status_output = gr.Textbox(label="Run Status / Submission Result", lines=5, interactive=False)
+    # Removed max_rows=10 from DataFrame constructor
+    results_table = gr.DataFrame(label="Questions and Agent Answers", wrap=True)
+
+    run_button.click(
+        fn=run_and_submit_all,
+        outputs=[status_output, results_table]
+    )
+
+if __name__ == "__main__":
+    # print("LangGraph version:", langgraph.__version__) 
+    print("\n" + "-"*30 + " App Starting " + "-"*30)
+    # Check for SPACE_HOST and SPACE_ID at startup for information
+    space_host_startup = os.getenv("SPACE_HOST")
+    space_id_startup = os.getenv("SPACE_ID") # Get SPACE_ID at startup
+    # import langgraph
+    # print("▶︎ LangGraph version:", langgraph.__version__)
+    if space_host_startup:
+        print(f"✅ SPACE_HOST found: {space_host_startup}")
+        print(f"   Runtime URL should be: https://{space_host_startup}.hf.space")
+    else:
+        print("ℹ️  SPACE_HOST environment variable not found (running locally?).")
+
+    if space_id_startup: # Print repo URLs if SPACE_ID is found
+        print(f"✅ SPACE_ID found: {space_id_startup}")
+        print(f"   Repo URL: https://huggingface.co/spaces/{space_id_startup}")
+        print(f"   Repo Tree URL: https://huggingface.co/spaces/{space_id_startup}/tree/main")
+    else:
+        print("ℹ️  SPACE_ID environment variable not found (running locally?). Repo URL cannot be determined.")
+
+    print("-"*(60 + len(" App Starting ")) + "\n")
+
+    print("Launching Gradio Interface for Basic Agent Evaluation...")
+    demo.launch(debug=True, share=False)