Pass system message as system_instruction to Gemini #1011.

.secrets.baseline

@@ -1401,16 +1401,6 @@
         "is_secret": false
       }
     ],
-    "website/docs/user-guide/models/google-gemini.mdx": [
-      {
-        "type": "Secret Keyword",
-        "filename": "website/docs/user-guide/models/google-gemini.mdx",
-        "hashed_secret": "026eec59a93703a5c95624645476e0d2a97cc84a",
-        "is_verified": false,
-        "line_number": 54,
-        "is_secret": false
-      }
-    ],
     "website/docs/user-guide/models/groq.mdx": [
       {
         "type": "Secret Keyword",
@@ -1578,5 +1568,5 @@
       }
     ]
   },
-  "generated_at": "2025-02-25T12:38:52Z"
+  "generated_at": "2025-02-26T21:26:48Z"
 }

autogen/oai/gemini.py

@@ -228,7 +228,7 @@ def create(self, params: dict) -> ChatCompletion:
         messages = params.get("messages", [])
         stream = params.get("stream", False)
         n_response = params.get("n", 1)
-        system_instruction = params.get("system_instruction")
+        system_instruction = self._extract_system_instruction(messages)
         response_validation = params.get("response_validation", True)
         tools = self._tools_to_gemini_tools(params["tools"]) if "tools" in params else None
 
@@ -387,6 +387,21 @@ def create(self, params: dict) -> ChatCompletion:
 
         return response_oai
 
+    def _extract_system_instruction(self, messages: list[dict]) -> str | None:
+        """Extract system instruction if provided."""
+        if messages is None or len(messages) == 0 or messages[0].get("role") != "system":
+            return None
+
+        message = messages.pop(0)
+        content = message["content"]
+
+        # Multi-model uses a list of dictionaries as content with text for the system message
+        # Otherwise normal agents will have strings as content
+        content = content[0].get("text", "").strip() if isinstance(content, list) else content.strip()
+
+        content = content if len(content) > 0 else None
+        return content
+
     def _oai_content_to_gemini_content(self, message: dict[str, Any]) -> tuple[list[Any], str]:
         """Convert AG2 content to Gemini parts, catering for text and tool calls"""
         rst = []
@@ -816,7 +831,13 @@ def total_cost_k(cost_per_k_input: float, cost_per_k_output: float):
         # Vertex AI pricing - based on Text input
         # https://cloud.google.com/vertex-ai/generative-ai/pricing#vertex-ai-pricing
 
-        if "gemini-1.5-flash" in model_name:
+        if "gemini-2.0-flash-lite" in model_name:
+            return total_cost_mil(0.075, 0.3)
+
+        elif "gemini-2.0-flash" in model_name:
+            return total_cost_mil(0.15, 0.6)
+
+        elif "gemini-1.5-flash" in model_name:
             if up_to_128k:
                 return total_cost_k(0.00001875, 0.000075)
             else:
@@ -841,7 +862,15 @@ def total_cost_k(cost_per_k_input: float, cost_per_k_output: float):
     else:
         # Non-Vertex AI pricing
 
-        if "gemini-1.5-flash-8b" in model_name:
+        if "gemini-2.0-flash-lite" in model_name:
+            # https://ai.google.dev/gemini-api/docs/pricing#gemini-2.0-flash-lite
+            return total_cost_mil(0.075, 0.3)
+
+        elif "gemini-2.0-flash" in model_name:
+            # https://ai.google.dev/gemini-api/docs/pricing#gemini-2.0-flash
+            return total_cost_mil(0.1, 0.4)
+
+        elif "gemini-1.5-flash-8b" in model_name:
             # https://ai.google.dev/pricing#1_5flash-8B
             if up_to_128k:
                 return total_cost_mil(0.0375, 0.15)

test/oai/test_gemini.py

@@ -91,6 +91,33 @@ def test_vertexai_initialization(self):
         assert vertexai_global_config.project == "fake-project-id", "Incorrect VertexAI project initialization"
         assert vertexai_global_config.credentials == mock_credentials, "Incorrect VertexAI credentials initialization"
 
+    def test_extract_system_instruction(self, gemini_client):
+        # Test: valid system instruction
+        messages = [{"role": "system", "content": "You are my personal assistant."}]
+        assert gemini_client._extract_system_instruction(messages) == "You are my personal assistant."
+
+        # Test: empty system instruction
+        messages = [{"role": "system", "content": " "}]
+        assert gemini_client._extract_system_instruction(messages) is None
+
+        # Test: the first message is not a system instruction
+        messages = [
+            {"role": "user", "content": "Hello!"},
+            {"role": "system", "content": "You are my personal assistant."},
+        ]
+        assert gemini_client._extract_system_instruction(messages) is None
+
+        # Test: empty message list
+        assert gemini_client._extract_system_instruction([]) is None
+
+        # Test: None input
+        assert gemini_client._extract_system_instruction(None) is None
+
+        # Test: system message without "content" key
+        messages = [{"role": "system"}]
+        with pytest.raises(KeyError):
+            gemini_client._extract_system_instruction(messages)
+
     def test_gemini_message_handling(self, gemini_client):
         messages = [
             {"role": "system", "content": "You are my personal assistant."},