Update app.py

app.py

@@ -16,16 +16,41 @@ from huggingface_hub import InferenceClient
 
 DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
 
+
 # ================================
 #   FUNÇÕES AUXILIARES
 # ================================
 
 def clean_answer(text: str) -> str:
+    """
+    Limpa a resposta do modelo para bater em EXACT MATCH:
+    - remove blocos <think>...</think> (Qwen Thinking)
+    - remove tags <think> soltas
+    - remove tags HTML genéricas
+    - remove prefixos tipo 'Final answer', 'Answer:'
+    - remove aspas externas
+    - normaliza espaços e ponto final solto
+    """
     if not text:
         return ""
 
     text = str(text).strip()
 
+    # Remover blocos <think>...</think>
+    text = re.sub(
+        r"<think>.*?</think>",
+        "",
+        text,
+        flags=re.DOTALL | re.IGNORECASE
+    ).strip()
+
+    # Remover tags <think> / </think> soltas
+    text = re.sub(r"</?think>", "", text, flags=re.IGNORECASE).strip()
+
+    # Remover qualquer tag HTML genérica
+    text = re.sub(r"<[^>]+>", "", text).strip()
+
+    # Remover prefixos do tipo "Final answer", "Answer:", etc.
     patterns_to_remove = [
         r"(?i)^final answer[:\- ]*",
         r"(?i)^answer[:\- ]*",
@@ -35,40 +60,71 @@ def clean_answer(text: str) -> str:
     for p in patterns_to_remove:
         text = re.sub(p, "", text).strip()
 
-    text = text.replace("\n", " ").replace("\r", " ").strip()
-    text = re.sub(r"\s+", " ", text).strip()
+    # Remover aspas externas
+    if len(text) > 2 and text.startswith('"') and text.endswith('"'):
+        text = text[1:-1].strip()
+    if len(text) > 2 and text.startswith("'") and text.endswith("'"):
+        text = text[1:-1].strip()
 
-    if len(text) > 2 and text.startswith(("'", '"')) and text.endswith(("'", '"')):
-        text = text[1:-1]
+    # Normalizar espaços
+    text = re.sub(r"\s+", " ", text).strip()
 
+    # Tirar ponto final solto
     if text.endswith(".") and not re.search(r"[0-9A-Za-z][.!?]$", text[:-1]):
-        text = text[:-1]
+        text = text[:-1].strip()
 
-    return text.strip()
+    return text
 
 
 def enforce_numeric_format(question: str, answer: str) -> str:
+    """
+    Pós-processa a resposta para:
+    - garantir duas casas decimais quando pedido
+    - extrair inteiros quando a pergunta é "how many / number of / what year"
+    - extrair códigos (NASA award, IOC code, etc.) quando a pergunta pede isso
+    """
     q = question.lower()
+    a = answer
 
+    # 1) Valores com duas casas decimais (ex: USD)
     if "two decimal places" in q or "2 decimal places" in q:
-        match = re.search(r"[-+]?\d+(?:[.,]\d+)?", answer)
+        match = re.search(r"[-+]?\d+(?:[.,]\d+)?", a)
         if match:
             try:
                 value = float(match.group(0).replace(",", ""))
                 return f"{value:.2f}"
-            except:
+            except Exception:
                 pass
 
+    # 2) Perguntas tipo "how many", "number of", "what year", "in which year"
     if any(kw in q for kw in ["how many", "number of", "what year", "in which year"]):
-        match = re.search(r"-?\d+", answer.replace(",", ""))
+        match = re.search(r"-?\d+", a.replace(",", ""))
         if match:
             return match.group(0)
 
-    return answer
+    # 3) Códigos tipo "IOC country code", "award number", "NASA award"
+    if (
+        "ioc country code" in q
+        or "award number" in q
+        or "nasa award" in q
+        or "grant number" in q
+        or "award no." in q
+    ):
+        # Procura tokens alfanuméricos em MAIÚSCULAS (3+ chars)
+        tokens = re.findall(r"[A-Z0-9]{3,}", a)
+        if tokens:
+            # Heurística simples: pega o token mais longo
+            best = max(tokens, key=len)
+            return best
+
+    return a
 
 
 def web_search(question: str, max_results: int = 5) -> str:
-    snippets = []
+    """
+    Usa DuckDuckGo (ddgs) pra buscar snippets de contexto.
+    """
+    snippets: List[str] = []
     try:
         with DDGS() as ddgs:
             for r in ddgs.text(question, max_results=max_results, safesearch="moderate"):
@@ -87,6 +143,9 @@ def web_search(question: str, max_results: int = 5) -> str:
 
 
 def get_file_context(api_url: str, task_id: str, item: dict) -> str:
+    """
+    Tenta baixar o arquivo de /files/{task_id} e extrair texto/planilha.
+    """
     file_name = (
         item.get("file_name")
         or item.get("filename")
@@ -114,7 +173,7 @@ def get_file_context(api_url: str, task_id: str, item: dict) -> str:
         if any(name_lower.endswith(ext) for ext in [".txt", ".csv", ".tsv"]):
             try:
                 text = data.decode("utf-8", errors="replace")
-            except:
+            except Exception:
                 text = data.decode("latin-1", errors="replace")
             return f"[FILE TXT]\n{text[:8000]}"
 
@@ -128,7 +187,8 @@ def get_file_context(api_url: str, task_id: str, item: dict) -> str:
                 print("[EXCEL PARSE ERROR]", e)
                 return "[FILE] Spreadsheet exists but cannot parse."
 
-        return f"[FILE BINARY: {file_name}] {len(data)} bytes"
+        # Outros tipos
+        return f"[FILE BINARY: {file_name}] {len(data)} bytes (type: {content_type})"
 
     except Exception as e:
         print("[FILE ERROR]", e)
@@ -254,7 +314,13 @@ def run_and_submit_all(profile: Optional[gr.OAuthProfile]):
         answer = agent(qtext, file_context)
 
         answers_payload.append({"task_id": qid, "submitted_answer": answer})
-        results_log.append({"Task ID": qid, "Question": qtext, "Submitted Answer": answer})
+        results_log.append(
+            {
+                "Task ID": qid,
+                "Question": qtext,
+                "Submitted Answer": answer,
+            }
+        )
 
     submission = {
         "username": username,
@@ -298,4 +364,4 @@ with gr.Blocks() as demo:
 
 
 if __name__ == "__main__":
-    demo.launch(debug=True, share=False)
+    demo.launch(debug=True, share=False)