Default to Gemini 2.5-flash (thinking disabled) for text/judge nodes

Key fixed on a non-billing project; 2.5-flash with thinking_budget=0 is fast and accurate (Mercedes Sosa -> correct '3'). Pace lowered to 0.15 rps for the ~10 RPM free tier.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

.env.example

@@ -4,15 +4,14 @@ TAVILY_API_KEY=
 # Required when LLM_PROVIDER=gemini (default). Free key: aistudio.google.com
 GOOGLE_API_KEY=
 
-# --- LLM provider for text/reasoning + judge nodes: "groq" or "gemini" ---
-# Default "groq" + llama-3.1-8b-instant works with the existing key (high TPM).
-# Switch to "gemini" for better quality once you have a free-tier GOOGLE_API_KEY
-# (created in a Google project WITHOUT billing enabled, else free tier = 0).
-LLM_PROVIDER=groq
+# --- LLM provider for text/reasoning + judge nodes: "gemini" or "groq" ---
+# Default "gemini" + gemini-2.5-flash (best quality, high free-tier TPM).
+# GOOGLE_API_KEY must come from a project WITHOUT billing (else free tier = 0).
+LLM_PROVIDER=gemini
+GEMINI_TEXT_MODEL=gemini-2.5-flash
 GROQ_TEXT_MODEL=llama-3.1-8b-instant
-GEMINI_TEXT_MODEL=gemini-2.0-flash
-# Global pace across all text-LLM calls (req/sec). 0.2 = ~12/min.
-RATE_LIMIT_RPS=0.2
+# Global pace across all text-LLM calls (req/sec). 0.15 = ~9/min (<10 RPM free).
+RATE_LIMIT_RPS=0.15
 
 # --- Optional: LangSmith tracing (nice for debugging the graph) ---
 # Set TRACING=true and provide the API key to stream runs to smith.langchain.com.

gaia_agent/config.py

@@ -24,21 +24,19 @@ class Settings(BaseSettings):
     google_api_key: str = ""
 
     # --- LLM provider for the text/reasoning + judge nodes ---
-    # "groq" (default, works with the existing key) or "gemini" (better quality,
-    # needs a free-tier GOOGLE_API_KEY from a project WITHOUT billing enabled).
+    # "gemini" (default, best quality + high free-tier TPM) or "groq".
     # Vision + Whisper always use Groq.
-    llm_provider: str = "groq"
+    llm_provider: str = "gemini"
 
     # --- Models (override via env) ---
-    gemini_text_model: str = "gemini-2.0-flash"
-    # 8b-instant has ~30k TPM free (vs 70B's ~12k) so it survives the batch.
+    gemini_text_model: str = "gemini-2.5-flash"
     groq_text_model: str = "llama-3.1-8b-instant"
     groq_vision_model: str = "meta-llama/llama-4-scout-17b-16e-instruct"
     groq_whisper_model: str = "whisper-large-v3"
 
     # --- Rate limiting (requests/sec across all text-LLM calls) ---
-    # Gemini 2.0-flash free tier ~15 RPM; 0.2 rps = ~12/min keeps headroom.
-    rate_limit_rps: float = 0.2
+    # Gemini 2.5-flash free tier ~10 RPM; 0.15 rps = ~9/min keeps headroom.
+    rate_limit_rps: float = 0.15
 
     # --- API + control knobs ---
     gaia_api_url: str = "https://agents-course-unit4-scoring.hf.space"

gaia_agent/llm.py

@@ -45,12 +45,15 @@ def get_text_llm(temperature: float = 0.0):
     # Default: Gemini (high free-tier TPM).
     from langchain_google_genai import ChatGoogleGenerativeAI
 
+    # thinking_budget=0 disables 2.5-flash's reasoning tokens -> much faster and
+    # cheaper, which matters under the slow free-tier RPM pacing.
     return ChatGoogleGenerativeAI(
         model=s.gemini_text_model,
         google_api_key=s.google_api_key,
         temperature=temperature,
         max_retries=3,
         rate_limiter=limiter,
+        thinking_budget=0,
     )