feat: Reorder `smart_invoke` fallback to OpenRouter-Gemini-Groq, add adaptive Gemini model selection, and persist API tier for subsequent calls.

.gitignore

@@ -1,2 +1,4 @@
 .env
-.cursorignore
+.cursorignore
+.venv_old
+.venv

agent.py

@@ -47,6 +47,7 @@ load_dotenv()
 #     huggingfacehub_api_token=os.getenv("HUGGINGFACEHUB_API_TOKEN"),
 # )
 
+# Groq Model (Primary)
 model = ChatGroq(
     model="meta-llama/llama-4-scout-17b-16e-instruct",
     temperature=0,
@@ -70,6 +71,23 @@ gemini_model = ChatGoogleGenerativeAI(
     temperature=0,
 )
 
+# NVIDIA Model (Secondary Fallback)
+nvidia_model = ChatOpenAI(
+    model="nvidia/llama-3.1-405b-instruct",
+    openai_api_key=os.getenv("NVIDIA_API_KEY"),
+    openai_api_base="https://integrate.api.nvidia.com/v1",
+    temperature=0,
+)
+
+# Vercel Model (Tertiary Fallback)
+# Note: Adjust model and base_url if using a specific Vercel AI Gateway setup
+vercel_model = ChatOpenAI(
+    model="meta-llama/llama-3.3-70b-instruct",
+    openai_api_key=os.getenv("VERCEL_API_KEY"),
+    openai_api_base="https://gateway.ai.vercel.com/v1",
+    temperature=0,
+)
+
 def smart_invoke(msgs, use_tools=False, start_tier=0):
     """
     Tiered fallback: OpenRouter -> Gemini -> Groq.
@@ -78,14 +96,18 @@ def smart_invoke(msgs, use_tools=False, start_tier=0):
     primary = model_with_tools if use_tools else model
     secondary = openrouter_with_tools if use_tools else openrouter_model
     tertiary = gemini_with_tools if use_tools else gemini_model
+    quaternary = nvidia_with_tools if use_tools else nvidia_model
+    quinary = vercel_with_tools if use_tools else vercel_model
     
     # Adaptive Gemini names to try if 1.5 flash is 404
-    gemini_alternatives = ["gemini-2.5-flash", "gemini-2.5-flash-lite", "gemini-3.1-flash-lite", "gemini-3-flash"]
+    gemini_alternatives = ["gemini-2.5-flash-lite", "gemma-3-1b", "gemini-3-flash", "gemini-3.1-flash-lite"]
     
     tiers = [
         {"name": "OpenRouter", "model": secondary, "key": "OPENROUTER_API_KEY"},
         {"name": "Gemini", "model": tertiary, "key": "GOOGLE_API_KEY", "alternatives": gemini_alternatives},
         {"name": "Groq", "model": primary, "key": "GROQ_API_KEY"},
+        {"name": "NVIDIA", "model": quaternary, "key": "NVIDIA_API_KEY"},
+        {"name": "Vercel", "model": quinary, "key": "VERCEL_API_KEY"},
     ]
     
     last_exception = None
@@ -410,6 +432,8 @@ tools_by_name = {tool.name: tool for tool in tools}
 model_with_tools = model.bind_tools(tools)
 openrouter_with_tools = openrouter_model.bind_tools(tools)
 gemini_with_tools = gemini_model.bind_tools(tools)
+nvidia_with_tools = nvidia_model.bind_tools(tools)
+vercel_with_tools = vercel_model.bind_tools(tools)
 
 def answer_message(state: AgentState) -> AgentState:
     messages = state["messages"]

verify_fallback.py

@@ -0,0 +1,68 @@
+import os
+import sys
+from unittest.mock import MagicMock, patch
+
+# Mocking modules that might not be available or needed for this test
+sys.modules['cv2'] = MagicMock()
+sys.modules['whisper'] = MagicMock()
+
+# Set dummy env vars BEFORE importing agent
+os.environ["OPENROUTER_API_KEY"] = "dummy"
+os.environ["GOOGLE_API_KEY"] = "dummy"
+os.environ["GROQ_API_KEY"] = "dummy"
+os.environ["NVIDIA_API_KEY"] = "dummy"
+os.environ["VERCEL_API_KEY"] = "dummy"
+
+# Add the current directory to path so we can import agent
+sys.path.append(os.getcwd())
+
+import agent
+from langchain_core.messages import HumanMessage
+
+def test_fallback_logic():
+    print("Testing fallback logic...")
+    
+    # Mock the invoke method for each tier's model
+    # Tiers: 0:OpenRouter, 1:Gemini, 2:Groq, 3:NVIDIA, 4:Vercel
+    
+    with patch('agent.openrouter_model.invoke') as mock_openrouter, \
+         patch('agent.gemini_model.invoke') as mock_gemini, \
+         patch('agent.model.invoke') as mock_groq, \
+         patch('agent.nvidia_model.invoke') as mock_nvidia, \
+         patch('agent.vercel_model.invoke') as mock_vercel:
+        
+        # Simulate failure for all tiers up to NVIDIA
+        mock_openrouter.side_effect = Exception("Rate limit (429)")
+        mock_gemini.side_effect = Exception("Rate limit (429)")
+        mock_groq.side_effect = Exception("Rate limit (429)")
+        
+        # NVIDIA should succeed
+        mock_nvidia.return_value = MagicMock(content="NVIDIA response")
+        
+        msgs = [HumanMessage(content="Hello")]
+        response, tier_idx = agent.smart_invoke(msgs, use_tools=False)
+        
+        print(f"Response from tier {tier_idx}: {response.content}")
+        assert tier_idx == 3
+        assert response.content == "NVIDIA response"
+        print("Fallback to NVIDIA successful!")
+
+        # Now simulate failure up to Vercel
+        mock_nvidia.side_effect = Exception("Rate limit (429)")
+        mock_vercel.return_value = MagicMock(content="Vercel response")
+        
+        response, tier_idx = agent.smart_invoke(msgs, use_tools=False)
+        print(f"Response from tier {tier_idx}: {response.content}")
+        assert tier_idx == 4
+        assert response.content == "Vercel response"
+        print("Fallback to Vercel successful!")
+
+if __name__ == "__main__":
+    try:
+        test_fallback_logic()
+        print("All fallback tests passed!")
+    except Exception as e:
+        print(f"Test failed: {e}")
+        import traceback
+        traceback.print_exc()
+        sys.exit(1)

verify_simple.py

@@ -0,0 +1,36 @@
+import os
+import sys
+from unittest.mock import MagicMock
+
+# Mocking modules
+sys.modules['cv2'] = MagicMock()
+sys.modules['whisper'] = MagicMock()
+
+# Set dummy env vars
+os.environ["OPENROUTER_API_KEY"] = "dummy"
+os.environ["GOOGLE_API_KEY"] = "dummy"
+os.environ["GROQ_API_KEY"] = "dummy"
+os.environ["NVIDIA_API_KEY"] = "dummy"
+os.environ["VERCEL_API_KEY"] = "dummy"
+
+sys.path.append(os.getcwd())
+
+import agent
+
+def verify_tiers():
+    from langchain_core.messages import HumanMessage
+    
+    # We can't easily call smart_invoke without real models unless we mock heavily.
+    # Let's just check the tiers list structure in a dummy call.
+    # Actually, we can't easily access 'tiers' inside smart_invoke as it's a local variable.
+    
+    # Let's check the global model objects.
+    print(f"NVIDIA model initialized: {agent.nvidia_model is not None}")
+    print(f"Vercel model initialized: {agent.vercel_model is not None}")
+    
+    # Check if they have invoke (they should)
+    print(f"NVIDIA model hasattr invoke: {hasattr(agent.nvidia_model, 'invoke')}")
+    print(f"Vercel model hasattr invoke: {hasattr(agent.vercel_model, 'invoke')}")
+
+if __name__ == "__main__":
+    verify_tiers()