commit

app.py

@@ -45,7 +45,7 @@ def clean_answer(text: str) -> str:
 
     if s.lower() in ["yes", "no"]:
         s = s.capitalize()
-        
+
     return s
 
 def classify_question(q: str) -> str:
@@ -89,30 +89,37 @@ class BasicAgent:
         print(f"Question: {question[:80]}...")
 
         qtype = classify_question(question)
-        rule = RULES[qtype]
 
-        # 검색 쿼리 보정 (숫자/날짜는 Wikipedia 우선)
-        search_query = question
-        if qtype in ["number", "date"]:
-            search_query = question + " site:wikipedia.org"
+        # 🔹 1단계: 검색 없이 먼저 답 시도
+        base_prompt = f"""
+    You are solving a GAIA benchmark question.
+
+    Rules:
+    - Answer with ONLY the final answer
+    - No explanation
+    - No extra text
 
-        # 검색 (실패 시 1회 재시도)
+    Question:
+    {question}
+    """.strip()
+
+        base_response = llm.invoke([HumanMessage(content=base_prompt)])
+        base_answer = clean_answer(base_response.content)
+
+        # 🔹 숫자/연도/YesNo가 명확하면 바로 반환
+        if qtype in ["number", "date", "yesno"] and base_answer:
+            print(f"Answer (no search): {base_answer}")
+            return base_answer
+
+        # 🔹 2단계: 그래도 애매하면 검색 사용
         try:
-            search_result = search_tool.run(search_query)
-            if not search_result:
-                raise RuntimeError("empty search")
+            search_result = search_tool.run(question)
         except Exception:
-            try:
-                search_result = search_tool.run(question)
-            except Exception:
-                search_result = ""
+            search_result = ""
 
         prompt = f"""
     {SYSTEM_PROMPT}
 
-    Additional Rules:
-    - {rule}
-
     Question:
     {question}
 
@@ -123,7 +130,7 @@ class BasicAgent:
         response = llm.invoke([HumanMessage(content=prompt)])
         answer = clean_answer(response.content)
 
-        print(f"Answer: {answer}")
+        print(f"Answer (with search): {answer}")
         return answer
 
 def run_and_submit_all( profile: gr.OAuthProfile | None):