Switch model backend from Groq to Cerebras

Groq's free-tier TPM (6000-12000) was too small for this agent's per-call cost (2400-9000+ tokens). Cerebras free tier offers 30,000 TPM / 1,000,000 TPD - several times more headroom. Uses CEREBRAS_API_KEY, model cerebras/gpt-oss-120b.

app.py

@@ -46,11 +46,11 @@ class TokenPacer:
     Step_callback that tracks actual token usage per step in a trailing 60s
     window and sleeps as needed to stay under a tokens-per-minute budget.
     Necessary because a single call's fixed overhead (system prompt + tool
-    schemas + question, before any tool output) already runs 2,400-4,000
-    tokens on this agent - so llama-3.1-8b-instant's free-tier TPM cap (6000)
-    only fits ~1-2 calls per minute regardless of request frequency, which
-    smolagents' native requests_per_minute throttle can't account for since
-    it only paces call count, not size.
+    schemas + question, before any tool output) already runs 2,400+ tokens
+    on this agent and grows with context - smolagents' native
+    requests_per_minute throttle can't account for that since it only paces
+    call count, not size. This is a secondary safeguard: Cerebras' free-tier
+    TPM budget (30,000) is generous enough that this should rarely trigger.
     """
     def __init__(self, tokens_per_minute_budget: int = 5000):
         self.tokens_per_minute_budget = tokens_per_minute_budget
@@ -99,25 +99,21 @@ class MemoryTrimmer:
 # ----- THIS IS WERE YOU CAN BUILD WHAT YOU WANT ------
 class BasicAgent:
     def __init__(self):
-        model_id = os.getenv("AGENT_MODEL_ID", "groq/llama-3.1-8b-instant")
-        api_key = os.getenv("GROQ_API_KEY")
+        model_id = os.getenv("AGENT_MODEL_ID", "cerebras/gpt-oss-120b")
+        api_key = os.getenv("CEREBRAS_API_KEY")
         if not api_key:
-            print("Warning: GROQ_API_KEY is not set - the agent will fail to call the model.")
-
-        # requests_per_minute: smolagents' own ApiModel throttles calls at this
-        # rate before ever hitting the API (a bit under Groq's 30 RPM cap).
-        # retry=False: smolagents' default retries a failed call several
-        # times internally (invisible to us - confirmed via Groq's dashboard,
-        # 429s with sub-second gaps between requests). Only ~30% of GAIA
-        # questions need to succeed here, so a call that still gets rate
-        # limited (e.g. on tokens/minute, which this doesn't track) should
-        # just fail fast - the per-question try/except below skips it and
-        # moves on, which is much cheaper than retry-storming.
+            print("Warning: CEREBRAS_API_KEY is not set - the agent will fail to call the model.")
+
+        # Cerebras' free tier: 5 RPM, 30,000 TPM, 1,000,000 TPD - far more
+        # TPM headroom than Groq's free tier (6000-12000), which was the
+        # actual bottleneck there. requests_per_minute paces call frequency
+        # natively; retry=False avoids smolagents' internal retry-on-error
+        # firing extra invisible HTTP calls on a rate-limited response.
         self.model = LiteLLMModel(
             model_id=model_id,
             api_key=api_key,
             temperature=0,
-            requests_per_minute=float(os.getenv("RATE_LIMIT_RPM", "25")),
+            requests_per_minute=float(os.getenv("RATE_LIMIT_RPM", "4.5")),
             retry=False,
         )
         self.agent = CodeAgent(
@@ -129,7 +125,7 @@ class BasicAgent:
             ],
             # add_base_tools=True would add a duplicate DuckDuckGo search tool -
             # every tool's schema is baked into the system prompt on every call,
-            # and that fixed overhead is what's blowing the 6000 TPM budget.
+            # keeping that overhead small helps stay well under the TPM budget.
             additional_authorized_imports=[
                 "pandas", "numpy", "math", "re", "json", "itertools",
                 "collections", "statistics", "datetime", "io", "openpyxl", "PIL",
@@ -137,7 +133,7 @@ class BasicAgent:
             max_steps=7,  # keep runs short: growing history otherwise blows the TPM budget by step ~6
             step_callbacks=[
                 MemoryTrimmer(),
-                TokenPacer(tokens_per_minute_budget=int(os.getenv("RATE_LIMIT_TOKENS_PER_MINUTE", "5000"))),
+                TokenPacer(tokens_per_minute_budget=int(os.getenv("RATE_LIMIT_TOKENS_PER_MINUTE", "25000"))),
             ],
         )
         print("BasicAgent initialized.")
@@ -311,7 +307,7 @@ with gr.Blocks() as demo:
         Once clicking on the "submit button, it can take quite some time ( this is the time for the agent to go through all the questions).
         This space provides a basic setup and is intentionally sub-optimal to encourage you to develop your own, more robust solution. For instance for the delay process of the submit button, a solution could be to cache the answers and submit in a seperate action or even to answer the questions in async.
 
-        **Setup:** This agent calls Groq (Llama 3.1 8B Instant) via `smolagents`. Get a free key at https://console.groq.com/keys and set it as the `GROQ_API_KEY` secret in this Space's settings before running.
+        **Setup:** This agent calls Cerebras (gpt-oss-120b) via `smolagents`. Get a free key at https://cloud.cerebras.ai and set it as the `CEREBRAS_API_KEY` secret in this Space's settings before running.
         """
     )