Update app.py

app.py

@@ -1,6 +1,7 @@
 import os
 import re
 import base64
+import json
 from io import BytesIO
 from pathlib import Path
 from zipfile import ZipFile
@@ -369,6 +370,12 @@ class BasicAgent:
         gemini_api_key = os.getenv("GEMINI_API_KEY")
         configured_model = os.getenv("GAIA_MODEL_ID")
 
+        if not gemini_api_key:
+            raise RuntimeError(
+                "O secret GEMINI_API_KEY é obrigatório porque todas as "
+                "respostas passam pela revisão final do Gemini."
+            )
+
         model_id = configured_model or DEFAULT_HF_MODEL
 
         if model_id.startswith("gemini/"):
@@ -391,21 +398,44 @@ class BasicAgent:
             max_tokens=2_000,
         )
         self.hf_token = hf_token
+        self.gemini_api_key = gemini_api_key
         self.model_id = model_id
         print(f"Modelo principal selecionado: {model_id}")
+
+        web_search_tool = DuckDuckGoSearchTool(
+            max_results=8, rate_limit=1.0
+        )
+        web_search_tool.description = (
+            "Searches the public web and returns result titles, URLs, and short "
+            "snippets. Use it to discover candidate sources. It does NOT open "
+            "or read the full pages; call visit_webpage on a returned URL."
+        )
+        visit_page_tool = VisitWebpageTool(max_output_length=30_000)
+        visit_page_tool.description = (
+            "Opens one exact HTTP/HTTPS URL and returns the readable page "
+            "content. Use it after web_search when facts depend on the page "
+            "itself, a table, article text, or linked source."
+        )
+        wikipedia_tool = WikipediaSearchTool(
+            user_agent="GAIA-Course-Agent/1.0 (educational project)",
+            language="en",
+        )
+        wikipedia_tool.description = (
+            "Searches English Wikipedia content directly. Use it for questions "
+            "that explicitly mention Wikipedia or for encyclopedic facts. "
+            "For version-specific or nomination details, verify the exact page "
+            "or archive with visit_webpage."
+        )
+
         agent_tools = [
-            DuckDuckGoSearchTool(max_results=8, rate_limit=1.0),
-            VisitWebpageTool(max_output_length=30_000),
-            WikipediaSearchTool(
-                user_agent="GAIA-Course-Agent/1.0 (educational project)",
-                language="en",
-            ),
+            web_search_tool,
+            visit_page_tool,
+            wikipedia_tool,
             InspectGaiaAttachmentTool(),
             YouTubeTranscriptTool(),
             AnalyzeGaiaImageTool(),
+            ConsultGeminiTool(),
         ]
-        if gemini_api_key:
-            agent_tools.append(ConsultGeminiTool())
 
         self.agent = CodeAgent(
             tools=agent_tools,
@@ -430,28 +460,38 @@ class BasicAgent:
 
         exact_match_prompt = """
 You are an expert AI assistant solving tasks from the GAIA benchmark.
-Research carefully before answering and cross-check uncertain facts.
-Use web_search to find sources and visit_webpage to read a result in detail.
-If the task mentions an attached file, call inspect_gaia_attachment with the
-task_id given in the task.
-If the task asks what someone says in a YouTube video, call youtube_transcript
-with the exact video URL before searching the web.
-If the task depends on an attached image, call analyze_gaia_image with the
-task_id and complete question. Do not try to infer image contents from metadata.
-When consult_gemini is available, use it selectively for a second opinion after
-you have collected evidence, especially when sources conflict or your candidate
-answer is uncertain. Give it the complete question and relevant evidence. Do
-not blindly copy its response; compare it with the sources before deciding.
-Prefer primary or official sources. When search snippets conflict, open the
-source and verify the relevant passage instead of guessing.
-Only call tools that are explicitly available. Never invent a function such as
-visit_webpage if it is not listed, and never use subprocess or shell commands.
-Do not repeat nearly identical searches. If one approach fails, change source
-or method.
-Your final answer must contain only the exact requested answer. Do not add
-explanations, conversational text, Markdown, citations, or the words
-"FINAL ANSWER". If a comma-separated list is requested, return only that list.
-If a number is requested, return only that number.
+
+TOOL ROUTING POLICY:
+1. web_search discovers URLs and snippets. It does not read full pages.
+2. visit_webpage opens and reads one exact URL. Use it after web_search to
+   verify articles, tables, archives, papers, and linked primary sources.
+3. wikipedia_search searches English Wikipedia. Use it when Wikipedia is
+   explicitly mentioned or for encyclopedic facts. For revision-specific,
+   nomination, archive, or table details, verify the exact page.
+4. inspect_gaia_attachment reads the official file for the supplied task_id.
+   Call it first for attached PDFs, spreadsheets, documents, audio, or code.
+5. analyze_gaia_image reads image pixels. Use it for diagrams, screenshots,
+   chess positions, or questions that visually depend on an attached image.
+6. youtube_transcript retrieves spoken subtitles. Use it when asked what a
+   person said. It cannot answer purely visual video questions.
+7. consult_gemini gives a second opinion. Use it after gathering evidence when
+   sources conflict or the candidate answer is uncertain. Include the complete
+   question, evidence, and candidate answer.
+
+Research carefully, prefer primary or official sources, and cross-check
+uncertain facts. A search snippet alone is insufficient when the source page
+can be opened. Never invent a tool, use subprocess, or use shell commands.
+Do not repeat nearly identical searches; change the source or method.
+
+FINAL RESPONSE POLICY:
+Call final_answer with only the requested value. Never include reasoning,
+explanations, labels, Markdown, citations, or the words "FINAL ANSWER".
+- Quantity/count: return only the number, unless units or currency are requested.
+- Person: return only the requested name component.
+- City/country/code: return only that value.
+- List: return only items with the requested separator and ordering.
+- Chess move: return only algebraic notation.
+- Quote: return only the requested spoken words.
 """
         self.agent.prompt_templates["system_prompt"] = (
             exact_match_prompt.strip()
@@ -484,7 +524,12 @@ If a number is requested, return only that number.
                     "(GEMINI_API_KEY para Gemini ou HF_TOKEN para Hugging Face)."
                 ) from exc
             raise
-        return self.format_exact_answer(question, str(result))
+        candidate = self.format_exact_answer(question, str(result))
+        return self.review_answer_with_gemini(
+            question=question,
+            candidate=candidate,
+            task_id=task_id,
+        )
 
     @staticmethod
     def deterministic_answer_cleanup(answer: str) -> str:
@@ -548,6 +593,90 @@ If a number is requested, return only that number.
 
         return self.deterministic_answer_cleanup(cleaned)
 
+    def review_answer_with_gemini(
+        self, question: str, candidate: str, task_id: str | None = None
+    ) -> str:
+        """Revisa obrigatoriamente conteúdo e formato antes de salvar a resposta."""
+        reviewer_model = os.getenv(
+            "GAIA_GEMINI_REVIEW_MODEL", DEFAULT_GEMINI_MODEL
+        )
+        review_prompt = f"""
+You are the mandatory final reviewer for a GAIA exact-match answer.
+
+Task ID: {task_id or "test"}
+Question:
+{question}
+
+Candidate answer:
+{candidate}
+
+Review the candidate for both likely correctness and exact requested format.
+Preserve it unless there is a clear factual, logical, ordering, spelling, or
+formatting error. Never add explanations to final_answer.
+
+Formatting rules:
+- If asked "how many", for a count, or for a numeric output: final_answer must
+  contain only the number, unless the question explicitly requests currency,
+  units, decimals, or another representation.
+- If asked for a first name, surname, city, country, or IOC code: return only
+  that requested value.
+- If asked for a list: return only the list, with the exact requested separator,
+  ordering, capitalization, and plurality.
+- If asked for a quote: return only the requested spoken words.
+- If asked for a chess move: return only algebraic notation.
+- Never include labels, Markdown, citations, rationale, or "FINAL ANSWER".
+
+Return one JSON object and nothing else:
+{{
+  "final_answer": "exact value to submit",
+  "changed": true,
+  "review_note": "brief reason, maximum 20 words"
+}}
+""".strip()
+
+        last_error = None
+        for use_json_mode in (True, False):
+            try:
+                kwargs = {
+                    "model": reviewer_model,
+                    "api_key": self.gemini_api_key,
+                    "messages": [{"role": "user", "content": review_prompt}],
+                    "temperature": 0,
+                    "max_tokens": 350,
+                }
+                if use_json_mode:
+                    kwargs["response_format"] = {"type": "json_object"}
+                response = completion(**kwargs)
+                content = str(response.choices[0].message.content).strip()
+                content = re.sub(
+                    r"^```(?:json)?\s*|\s*```$", "", content, flags=re.I
+                )
+                start = content.find("{")
+                end = content.rfind("}")
+                if start >= 0 and end > start:
+                    content = content[start : end + 1]
+                payload = json.loads(content)
+                final_answer = self.deterministic_answer_cleanup(
+                    str(payload.get("final_answer", ""))
+                )
+                if not final_answer:
+                    raise ValueError("Gemini returned an empty final_answer.")
+
+                changed = final_answer != candidate
+                note = str(payload.get("review_note", "")).strip()
+                print(f"Candidate answer: {candidate}")
+                print(f"Gemini reviewed answer: {final_answer}")
+                print(f"Gemini changed answer: {changed}")
+                print(f"Gemini review note: {note}")
+                return final_answer
+            except Exception as exc:
+                last_error = exc
+
+        raise RuntimeError(
+            "A revisão obrigatória do Gemini falhou; a resposta não foi salva. "
+            f"Detalhe: {last_error}"
+        )
+
 
 def empty_results() -> pd.DataFrame:
     return pd.DataFrame(columns=RESULT_COLUMNS)