初始化

server/agent_service.py

@@ -0,0 +1,56 @@
+import asyncio
+
+from autogen_agentchat.agents import AssistantAgent
+from autogen_agentchat.messages import MultiModalMessage, TextMessage
+from autogen_core import Image
+from autogen_core.models import ModelFamily
+from autogen_ext.models.ollama import OllamaChatCompletionClient
+
+from . import config
+from .mcp_tools import load_mcp_tools
+
+
+class AvatarAgentService:
+    def __init__(self) -> None:
+        self._model_client = OllamaChatCompletionClient(
+            model=config.OLLAMA_MODEL,
+            model_info={
+                "vision": True,
+                "function_calling": True,
+                "json_output": True,
+                "family": ModelFamily.UNKNOWN,
+                "structured_output": True,
+            },
+        )
+        self._agent: AssistantAgent | None = None
+        self._agent_lock = asyncio.Lock()
+
+    async def _create_agent(self) -> AssistantAgent:
+        tools = await load_mcp_tools()
+        return AssistantAgent(
+            name="avatar",
+            model_client=self._model_client,
+            system_message=config.SYSTEM_MESSAGE,
+            tools=tools or None,
+            reflect_on_tool_use=bool(tools),
+        )
+
+    async def _get_agent(self) -> AssistantAgent:
+        if self._agent is not None:
+            return self._agent
+        async with self._agent_lock:
+            if self._agent is None:
+                self._agent = await self._create_agent()
+        return self._agent
+
+    async def reply(self, user_text: str, image_b64: str) -> str:
+        agent = await self._get_agent()
+        user_image = Image.from_base64(image_b64)
+        multimodal_task = MultiModalMessage(source="user", content=[user_text, user_image])
+
+        ai_response = ""
+        async for message in agent.run_stream(task=multimodal_task):
+            if isinstance(message, TextMessage) and message.source == "avatar":
+                ai_response = message.content
+
+        return ai_response