Update agent.py

agent.py

@@ -1,14 +1,17 @@
 """
-agent.py: free agent for the Hugging Face Agents Course, Unit 4 (GAIA level 1).
-
-Uses smolagents + a free Gemini API key (Google AI Studio) through LiteLLM.
-Put GEMINI_API_KEY in your Space: Settings -> Variables and secrets -> New secret.
-
-Faster version:
-- no fixed pause after every question: a shared rate limiter spaces out model calls
-  only as much as the free tier needs, so several questions can run in parallel
-- a fresh agent per question, so parallel runs don't share memory
-- fewer steps per question and short retries
+agent.py: free agent for the Hugging Face Agents Course, Unit 4 (GAIA level 1), using Groq.
+
+Setup:
+  1. Create a free key at console.groq.com -> API Keys (no credit card).
+  2. In your Space: Settings -> Variables and secrets -> New secret
+         name:  GROQ_API_KEY      value: your key
+  3. Restart the Space.
+
+How it stays fast and inside Groq's free limits:
+- a shared rate limiter spaces out model calls across the parallel questions
+- short retries when Groq answers "rate limit" (429)
+- a fresh agent per question, at most MAX_STEPS steps each
+- questions that need audio, video or images are skipped (Llama 3.3 can't use them)
 """
 
 import os
@@ -18,12 +21,11 @@ import tempfile
 import threading
 
 import requests
-from PIL import Image
 from smolagents import CodeAgent, LiteLLMModel, DuckDuckGoSearchTool, VisitWebpageTool
 
 API_URL = "https://agents-course-unit4-scoring.hf.space"   # same as DEFAULT_API_URL in app.py
-MODEL_ID = "groq/llama-3.3-70b-versatile"                         # any current free Gemini Flash model
-REQUESTS_PER_MINUTE = 9      # keep just under your free-tier limit (check it in Google AI Studio)
+MODEL_ID = "groq/llama-3.3-70b-versatile"                   # if Groq renames it, pick a current model at console.groq.com
+REQUESTS_PER_MINUTE = 10     # Groq's free plan also limits tokens per minute, so stay cautious
 MAX_STEPS = 6                # most level-1 questions need 2-5 steps
 
 INSTRUCTIONS = """You are answering a question from the GAIA benchmark. Your answer is checked by EXACT MATCH.
@@ -59,8 +61,8 @@ class RateLimiter:
 LIMITER = RateLimiter(REQUESTS_PER_MINUTE)
 
 
-class RateLimitedModel(LiteLLMModel):
-    """Gemini model that waits for the shared rate limiter and retries briefly on 429 errors."""
+class GroqModel(LiteLLMModel):
+    """Groq model that waits for the shared rate limiter and retries briefly on rate-limit errors."""
     def generate(self, *args, **kwargs):
         for attempt in range(4):
             LIMITER.wait()
@@ -68,21 +70,20 @@ class RateLimitedModel(LiteLLMModel):
                 return super().generate(*args, **kwargs)
             except Exception as e:
                 msg = str(e).lower()
-                if attempt < 3 and ("429" in msg or "rate" in msg or "quota" in msg or "overloaded" in msg):
-                    time.sleep(8 * (attempt + 1))
+                if attempt < 3 and ("429" in msg or "rate" in msg or "overloaded" in msg):
+                    time.sleep(10 * (attempt + 1))
                     continue
                 raise
 
 
+def _new_model():
+    return GroqModel(model_id=MODEL_ID, api_key=os.environ["GROQ_API_KEY"], temperature=0.0)
+
+
 def _make_agent():
-    model = RateLimitedModel(
-        model_id=MODEL_ID,
-        api_key=os.environ["GEMINI_API_KEY"],
-        temperature=0.0,
-    )
     return CodeAgent(
-        tools=[DuckDuckGoSearchTool(max_results=5), VisitWebpageTool(max_output_length=12000)],
-        model=model,
+        tools=[DuckDuckGoSearchTool(max_results=5), VisitWebpageTool(max_output_length=6000)],
+        model=_new_model(),
         add_base_tools=False,
         additional_authorized_imports=[
             "pandas", "numpy", "openpyxl", "json", "re", "math",
@@ -94,10 +95,11 @@ def _make_agent():
 
 
 def check_model():
-    """One tiny call to Gemini before the run. Returns None if it works, or the error text."""
+    """One tiny call to Groq before the run. Returns None if it works, or the error text."""
+    if not os.environ.get("GROQ_API_KEY"):
+        return "GROQ_API_KEY secret is missing. Add it in Settings -> Variables and secrets, then restart the Space."
     try:
-        model = LiteLLMModel(model_id=MODEL_ID, api_key=os.environ["GEMINI_API_KEY"], temperature=0.0)
-        model.generate([{"role": "user", "content": [{"type": "text", "text": "Reply with the word OK."}]}])
+        _new_model().generate([{"role": "user", "content": [{"type": "text", "text": "Reply with the word OK."}]}])
         return None
     except Exception as e:
         return f"{type(e).__name__}: {e}"
@@ -105,9 +107,6 @@ def check_model():
 
 class BasicAgent:
     def __init__(self):
-        if not os.environ.get("GEMINI_API_KEY"):
-            raise RuntimeError("GEMINI_API_KEY secret is missing (Space Settings -> Variables and secrets), "
-                               "or the Space wasn't restarted after adding it.")
         self.last_error = {}     # task_id -> error text, shown in the results table
 
     @staticmethod
@@ -137,24 +136,18 @@ class BasicAgent:
         return text.strip().rstrip(".").strip()
 
     def __call__(self, question, task_id=None):
-        prompt = INSTRUCTIONS + question
-        images = None
+        if "youtube.com" in question or "youtu.be" in question:
+            return "unknown"                               # video: can't be watched, skip fast
 
+        prompt = INSTRUCTIONS + question
         path = self._download_file(task_id)
         if path:
-            if path.lower().endswith((".mp3", ".wav", ".m4a")):
-                return "unknown"                           # audio can't be handled for free: skip fast
-            if path.lower().endswith((".png", ".jpg", ".jpeg", ".webp")):
-                images = [Image.open(path)]                # Gemini can look at images directly
-                prompt += "\n\nThe image attached to this question is provided to you."
-            else:
-                prompt += f"\n\nThe file for this question is saved at: {path}"
-
-        if "youtube.com" in question or "youtu.be" in question:
-            return "unknown"                               # video questions: skip instead of burning steps
+            if path.lower().endswith((".mp3", ".wav", ".m4a", ".png", ".jpg", ".jpeg", ".webp")):
+                return "unknown"                           # audio or image: Llama 3.3 can't use it, skip fast
+            prompt += f"\n\nThe file for this question is saved at: {path}"
 
         try:
-            answer = _make_agent().run(prompt, images=images)
+            answer = _make_agent().run(prompt)
         except Exception as e:
             err = f"{type(e).__name__}: {e}"
             print(f"[{task_id}] FAILED: {err}")