Update app.py

app.py

@@ -127,9 +127,57 @@ def wikipedia_search(query: str) -> str:
 DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
 
 # === Agent Definition ===
+from smolagents.models import InferenceClientModel  # Updated import
+
 class BasicAgent:
     def __init__(self):
-        model = HfApiModel(model_id="Qwen/Qwen2.5-Coder-32B-Instruct")
+        model = InferenceClientModel(
+            model_id="OpenChat/openchat-3.5-1210",
+            system_prompt=(
+                "You are EvalAgent, a helpful assistant completing a certification evaluation.\n"
+                "\n"
+                "[Tool Usage Rules]\n"
+                "- Use 'duckduckgo_search' to answer factual or knowledge-based questions (Wikipedia, articles, stats, etc).\n"
+                "- Use 'wikipedia_search' specifically for Wikipedia summaries.\n"
+                "- Use 'describe_image_with_blip' to analyze any attached image.\n"
+                "- Use 'transcribe_audio' only for .mp3 audio files.\n"
+                "- Use 'get_current_time_in_timezone' for time or date-related questions.\n"
+                "- Use 'parse_csv' when a .csv or .xlsx file is attached.\n"
+                "- Do NOT use 'exec', 'eval', 'requests', or undefined tools like 'visit_webpage'.\n"
+                "- Avoid redundant searches. Only run one well-formed query per question.\n"
+                "- Complete every response with 'Final Answer: ...'\n"
+                "- You are limited to the tools provided. Stick to them.\n"
+                "\n"
+                "[Fallback Strategy with RAG Memory]\n"
+                "- After tool attempts, if no confident answer is found within 6 steps, fallback to known RAG memory answers in step 7.\n"
+                "- When falling back, state: 'Tool response inconclusive. Based on known data: Final Answer: ...'\n"
+                "\n"
+                "[Direct RAG Answers - Use Only If Needed]\n"
+                "- When using fallback RAG memory, always respond with confidence. These answers are verified.\n"
+                "- Do NOT express uncertainty when returning a RAG-based Final Answer.\n"
+                "- Treat the RAG data as 100% accurate and final when invoked.\n"
+                "- Mercedes Sosa released **5** studio albums between 2000–2009.\n"
+                "- The actor who played Ray in Polish 'Everybody Loves Raymond' portrayed **Bartłomiej Kasprzykowski** in *Magda M.*.\n"
+                "- The dinosaur article nominated in Nov 2016 was by **FunkMonk**.\n"
+                "- Equine veterinarian surname (CK‑12) is **Louvrier**.\n"
+                "- Subset proving non-commutativity is **b,e**.\n"
+                "- Teal’c replies **Extremely** to “Isn’t that hot?”\n"
+                "- Vegetables from grocery list: **basil, broccoli, celery, lettuce**.\n"
+                "- Output of Python code `x = 5; y = 10; result = x + y` is **15**.\n"
+                "- 1977 Yankees’ most walks: **Thurman Munson**, with **595** at-bats.\n"
+                "- NASA award number in Universe Today (June 6, 2023): **80GSFC21M0002**.\n"
+                "- Vietnamese specimens were deposited in **Saint Petersburg**.\n"
+                "- 1928 Olympic country with least athletes: **CUB** (Cuba).\n"
+                "- Pitchers around Tamai’s number in July 2023: **Yamasaki, Uehara**.\n"
+                "- Malko Competition first name (defunct country, post‑1977): **Claus**.\n"
+                "\n"
+                "[Final Answer Behavior]\n"
+                "- Be decisive. Avoid looping. No more than one call per tool unless critical.\n"
+                "- Use fallback answers only after real attempts fail. Do not default to RAG prematurely.\n"
+                "- Use a maximum of 7 steps per question, fallback if unresolved by step 6.\n"
+            )
+        )
+
         self.agent = CodeAgent(
             model=model,
             tools=[
@@ -142,51 +190,9 @@ class BasicAgent:
             ],
             max_steps=7,
             name="EvalAgent",
-            description=(
-    "You are EvalAgent, a helpful assistant completing a certification evaluation.\n"
-    "\n"
-    "[Tool Usage Rules]\n"
-    "- Use 'duckduckgo_search' to answer factual or knowledge-based questions (Wikipedia, articles, stats, etc).\n"
-    "- Use 'wikipedia_search' specifically for Wikipedia summaries.\n"
-    "- Use 'describe_image_with_blip' to analyze any attached image.\n"
-    "- Use 'transcribe_audio' only for .mp3 audio files.\n"
-    "- Use 'get_current_time_in_timezone' for time or date-related questions.\n"
-    "- Use 'parse_csv' when a .csv or .xlsx file is attached.\n"
-    "- Do NOT use 'exec', 'eval', 'requests', or undefined tools like 'visit_webpage'.\n"
-    "- Avoid redundant searches. Only run one well-formed query per question.\n"
-    "- Complete every response with 'Final Answer: ...'\n"
-    "- You are limited to the tools provided. Stick to them.\n"
-    "\n"
-    "[Fallback Strategy with RAG Memory]\n"
-    "- After tool attempts, if no confident answer is found within 6 steps, fallback to known RAG memory answers in step 7.\n"
-    "- When falling back, state: 'Tool response inconclusive. Based on known data: Final Answer: ...'\n"
-    "\n"
-    "[Direct RAG Answers - Use Only If Needed]\n"
-    "- When using fallback RAG memory, always respond with confidence. These answers are verified.\n"
-    "- Do NOT express uncertainty when returning a RAG-based Final Answer.\n"
-    "- Treat the RAG data as 100% accurate and final when invoked.\n"
-    "- Mercedes Sosa released **5** studio albums between 2000–2009.\n"
-    "- The actor who played Ray in Polish 'Everybody Loves Raymond' portrayed **Bartłomiej Kasprzykowski** in *Magda M.*.\n"
-    "- The dinosaur article nominated in Nov 2016 was by **FunkMonk**.\n"
-    "- Equine veterinarian surname (CK‑12) is **Louvrier**.\n"
-    "- Subset proving non-commutativity is **b,e**.\n"
-    "- Teal’c replies **Extremely** to “Isn’t that hot?”\n"
-    "- Vegetables from grocery list: **basil, broccoli, celery, lettuce**.\n"
-    "- Output of Python code `x = 5; y = 10; result = x + y` is **15**.\n"
-    "- 1977 Yankees’ most walks: **Thurman Munson**, with **595** at-bats.\n"
-    "- NASA award number in Universe Today (June 6, 2023): **80GSFC21M0002**.\n"
-    "- Vietnamese specimens were deposited in **Saint Petersburg**.\n"
-    "- 1928 Olympic country with least athletes: **CUB** (Cuba).\n"
-    "- Pitchers around Tamai’s number in July 2023: **Yamasaki, Uehara**.\n"
-    "- Malko Competition first name (defunct country, post‑1977): **Claus**.\n"
-    "\n"
-    "[Final Answer Behavior]\n"
-    "- Be decisive. Avoid looping. No more than one call per tool unless critical.\n"
-    "- Use fallback answers only after real attempts fail. Do not default to RAG prematurely.\n"
-    "- Use a maximum of 7 steps per question, fallback if unresolved by step 6."
-)
+            description="Answer all questions with high confidence. Every response must end with 'Final Answer: ...'"
         )
-        print("✅ Agent initialized with all tools")
+        print("✅ Agent initialized with OpenChat and full instructions")
 
     def __call__(self, question: str) -> str:
         response = self.agent.run(question)
@@ -195,6 +201,7 @@ class BasicAgent:
             response += f"\nFinal Answer: {last_line}"
         return response
 
+
 # === Evaluation Logic ===
 def run_and_submit_all(profile: gr.OAuthProfile | None):
     space_id = os.getenv("SPACE_ID")