Update app.py

app.py

@@ -1,81 +1,55 @@
 import os
 import re
-import requests
 import traceback
+import requests
 import gradio as gr
 import pandas as pd
-import google.generativeai as genai
-from google.generativeai.types import Tool, GoogleSearch
+from google import genai
+from google.genai import types
 
 # --- Constants ---
 DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
 
 
-# ─────────────────────────────────────────────
-#  Helper: strip markdown / fences from output
-# ─────────────────────────────────────────────
 def clean_answer(text: str) -> str:
-    """
-    Normalise the model's raw output into a clean, exact-match-ready string.
-    """
     text = text.strip()
-
-    # Remove code fences if the whole reply is wrapped in one
     text = re.sub(r"^```[a-zA-Z]*\n?", "", text)
     text = re.sub(r"```$", "", text)
-
-    # Remove common label prefixes the model likes to add
     for prefix in [
         "Final Answer:", "Answer:", "FINAL ANSWER:", "ANSWER:",
         "The answer is:", "The final answer is:",
     ]:
         if text.lower().startswith(prefix.lower()):
             text = text[len(prefix):].strip()
-
-    # Collapse internal newlines to spaces, then trim
     text = " ".join(text.split())
-
-    # Do NOT split on "." — it breaks decimals, abbreviations, etc.
     return text[:200]
 
 
-# ─────────────────────────────────────────────
-#  Agent
-# ─────────────────────────────────────────────
 class BasicAgent:
-    """
-    A Gemini-powered agent that uses:
-      • Google Search grounding  — for real-time factual questions
-      • Code execution           — for maths / data problems
-      • Self-verification pass   — to catch obvious hallucinations
-
-    Falls back gracefully when tools aren't available.
-    """
-
-    SYSTEM_PROMPT = """You are an expert AI assistant solving GAIA benchmark questions.
+    SYSTEM = """You are an expert AI assistant solving GAIA benchmark questions.
 
 RULES:
 1. Think step-by-step before answering.
-2. Use Google Search when you need current facts, dates, or specific data.
-3. Use code execution for any arithmetic, unit conversions, or data processing.
-4. Return ONLY the final answer — no explanation, no preamble, no punctuation tail.
-5. For numbers: no commas, correct decimal places (e.g. 42 not 42.0 if whole number).
-6. For lists: comma-separated values in alphabetical order unless asked otherwise.
-7. For yes/no questions: answer exactly "yes" or "no" (lowercase).
-8. Keep the answer as short as possible while being complete and correct.
+2. Use Google Search for any factual, date, or real-world questions.
+3. Use code execution for arithmetic, unit conversions, or data processing.
+4. Return ONLY the final answer — no explanation, no preamble.
+5. Numbers: no commas, correct decimals (42 not 42.0 if whole).
+6. Lists: comma-separated, alphabetical unless told otherwise.
+7. Yes/no questions: answer exactly "yes" or "no" (lowercase).
+8. Be as concise as possible while being complete and correct.
 """
 
-    VERIFY_PROMPT = """Question: {question}
+    VERIFY = """Question: {question}
 
 Proposed answer: {answer}
 
-Review the answer carefully:
+Review:
 - Is it factually correct?
-- Is it in the exact format requested?
+- Is it in the exact requested format?
 - Is it as concise as possible?
 
 If correct, repeat it unchanged.
-If wrong or badly formatted, return ONLY the corrected answer.
+If wrong or mis-formatted, return ONLY the corrected answer.
 
 Return ONLY the final answer — nothing else."""
 
@@ -84,27 +58,25 @@ Return ONLY the final answer — nothing else."""
         if not api_key:
             raise EnvironmentError("GEMINI_API_KEY environment variable is not set.")
 
-        genai.configure(api_key=api_key)
+        self.client = genai.Client(api_key=api_key)
+        self.model_id = "gemini-2.5-flash-preview-04-17"
 
-        # Tools available in gemini-2.5-flash
-        self._search_tool = Tool(google_search=GoogleSearch())
+        self.tools = [
+            types.Tool(google_search=types.GoogleSearch()),
+            types.Tool(code_execution=types.ToolCodeExecution()),
+        ]
 
-        # Primary model — with search grounding + code execution
-        self.model = genai.GenerativeModel(
-            model_name="gemini-2.5-flash",
-            system_instruction=self.SYSTEM_PROMPT,
-            tools=[self._search_tool, "code_execution"],
+        self.main_config = types.GenerateContentConfig(
+            system_instruction=self.SYSTEM,
+            tools=self.tools,
         )
 
-        # Verification model — plain text, no tools (avoids circular loops)
-        self.verify_model = genai.GenerativeModel(
-            model_name="gemini-2.5-flash",
+        self.verify_config = types.GenerateContentConfig(
             system_instruction="You are a precise answer validator. Return ONLY the final answer.",
         )
 
-        print("✅ BasicAgent (Gemini 2.5 Flash + Search + Code) initialised.")
+        print("✅ BasicAgent (gemini-2.5-flash + Search + Code) initialised.")
 
-    # ------------------------------------------------------------------
     def __call__(self, question: str) -> str:
         try:
             return self._run(question)
@@ -112,52 +84,44 @@ Return ONLY the final answer — nothing else."""
             print(f"[AGENT ERROR] {exc}\n{traceback.format_exc()}")
             return "N/A"
 
-    # ------------------------------------------------------------------
     def _run(self, question: str) -> str:
-        # ── Pass 1: full reasoning with tools ──────────────────────────
-        response1 = self.model.generate_content(question)
-        ans1 = self._extract_text(response1)
+        resp1 = self.client.models.generate_content(
+            model=self.model_id,
+            contents=question,
+            config=self.main_config,
+        )
+        ans1 = clean_answer(self._extract_text(resp1))
+        print(f"  [Pass 1] {ans1!r}")
 
         if not ans1 or ans1 == "N/A":
             return "N/A"
 
-        ans1 = clean_answer(ans1)
-        print(f"  [Pass 1] {ans1!r}")
-
-        # ── Pass 2: self-verification (plain model, no tools) ──────────
-        verify_prompt = self.VERIFY_PROMPT.format(question=question, answer=ans1)
-        response2 = self.verify_model.generate_content(verify_prompt)
-        ans2 = clean_answer(self._extract_text(response2))
-
+        verify_prompt = self.VERIFY.format(question=question, answer=ans1)
+        resp2 = self.client.models.generate_content(
+            model=self.model_id,
+            contents=verify_prompt,
+            config=self.verify_config,
+        )
+        ans2 = clean_answer(self._extract_text(resp2))
         print(f"  [Pass 2] {ans2!r}")
 
         return ans2 if ans2 else ans1
 
-    # ------------------------------------------------------------------
     @staticmethod
     def _extract_text(response) -> str:
-        """
-        Pull plain text out of a GenerateContentResponse regardless of
-        whether it contains tool-use / code-execution blocks.
-        """
         try:
-            # Fast path — the .text property works when there's no ambiguity
-            return response.text.strip()
+            if response.text:
+                return response.text.strip()
         except Exception:
             pass
-
-        # Slow path — iterate parts
         texts = []
         for candidate in response.candidates:
             for part in candidate.content.parts:
                 if hasattr(part, "text") and part.text:
                     texts.append(part.text.strip())
-        return " ".join(texts).strip() if texts else "N/A"
+        return " ".join(texts).strip() or "N/A"
 
 
-# ─────────────────────────────────────────────
-#  Gradio runner
-# ─────────────────────────────────────────────
 def run_and_submit_all(profile: gr.OAuthProfile | None):
     space_id = os.getenv("SPACE_ID")
 
@@ -167,20 +131,16 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
     username = profile.username
     print(f"Logged in as: {username}")
 
-    api_url = DEFAULT_API_URL
-    questions_url = f"{api_url}/questions"
-    submit_url    = f"{api_url}/submit"
+    questions_url = f"{DEFAULT_API_URL}/questions"
+    submit_url    = f"{DEFAULT_API_URL}/submit"
 
-    # ── Instantiate agent ───────────────────────────────────────────────
     try:
         agent = BasicAgent()
     except Exception as e:
         return f"Error initialising agent: {e}", None
 
     agent_code = f"https://huggingface.co/spaces/{space_id}/tree/main" if space_id else "unknown"
-    print(f"Agent code URL: {agent_code}")
 
-    # ── Fetch questions ─────────────────────────────────────────────────
     print(f"Fetching questions from {questions_url} …")
     try:
         resp = requests.get(questions_url, timeout=15)
@@ -192,36 +152,28 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
     except Exception as e:
         return f"Error fetching questions: {e}", None
 
-    # ── Run agent ───────────────────────────────────────────────────────
-    results_log      = []
-    answers_payload  = []
+    results_log     = []
+    answers_payload = []
 
     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 malformed item: {item}")
             continue
 
-        print(f"\n[{task_id}] Q: {question_text[:120]}")
+        print(f"\n[{task_id}] {question_text[:120]}")
         try:
             answer = agent(question_text)
         except Exception as e:
             answer = f"AGENT ERROR: {e}"
-            print(f"  ERROR: {e}")
 
         print(f"  → {answer!r}")
         answers_payload.append({"task_id": task_id, "submitted_answer": answer})
-        results_log.append({
-            "Task ID": task_id,
-            "Question": question_text,
-            "Submitted Answer": answer,
-        })
+        results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": answer})
 
     if not answers_payload:
         return "Agent produced no answers.", pd.DataFrame(results_log)
 
-    # ── Submit ──────────────────────────────────────────────────────────
     submission_data = {
         "username":   username.strip(),
         "agent_code": agent_code,
@@ -240,8 +192,7 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
             f"Msg:   {result.get('message', '')}"
         )
     except requests.exceptions.HTTPError as e:
-        detail = e.response.text[:500]
-        final_status = f"❌ Submission failed (HTTP {e.response.status_code}): {detail}"
+        final_status = f"❌ HTTP {e.response.status_code}: {e.response.text[:500]}"
     except Exception as e:
         final_status = f"❌ Submission error: {e}"
 
@@ -249,23 +200,18 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
     return final_status, pd.DataFrame(results_log)
 
 
-# ─────────────────────────────────────────────
-#  Gradio UI
-# ─────────────────────────────────────────────
 with gr.Blocks() as demo:
     gr.Markdown("# GAIA Agent Evaluation Runner")
     gr.Markdown(
         """
         **How to use:**
-        1. Clone this Space and set your `GEMINI_API_KEY` secret.
+        1. Set `GEMINI_API_KEY` in your Space secrets.
         2. Log in with the button below.
-        3. Click **Run Evaluation** — the agent will answer all questions and submit automatically.
-
-        *Note: runs can take several minutes while the agent processes all questions.*
+        3. Click **Run Evaluation** — the agent answers all questions and submits automatically.
         """
     )
     gr.LoginButton()
-    run_btn = gr.Button("▶ Run Evaluation & Submit All Answers", variant="primary")
+    run_btn     = gr.Button("▶ Run Evaluation & Submit All Answers", variant="primary")
     status_out  = gr.Textbox(label="Status / Result", lines=6, interactive=False)
     results_tbl = gr.DataFrame(label="Questions & Agent Answers", wrap=True)