初始化提交

backend/app/ai_provider/routing/multimodal.py

@@ -0,0 +1,81 @@
+from __future__ import annotations
+
+from enum import StrEnum
+from typing import Any
+
+
+class MultiModalDecision(StrEnum):
+    """基于消息内容分析的选路决策。
+
+    每个值代表所选模型必须满足的一项能力需求。
+    PLAIN_TEXT 表示不需要特殊能力。
+    """
+
+    PLAIN_TEXT = "plain_text"
+    VISION = "vision"
+    THINKING = "thinking"
+    TOOL_USE = "tool_use"
+    STRUCTURED_OUTPUT = "structured_output"
+
+
+class MultiModalRouter:
+    """分析消息列表以确定所需的模型能力。
+
+    用法::
+
+        messages = [{"role": "user", "content": "Hello"}]
+        decisions = MultiModalRouter.decide_from_messages(messages)
+        # -> {MultiModalDecision.PLAIN_TEXT}
+
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "What's in this image?"},
+                    {"type": "image_url", "image_url": {"url": "..."}},
+                ],
+            },
+        ]
+        decisions = MultiModalRouter.decide_from_messages(messages)
+        # -> {MultiModalDecision.VISION}
+    """
+
+    @staticmethod
+    def decide_from_messages(messages: list[dict[str, Any]]) -> set[MultiModalDecision]:
+        """分析消息并返回所需能力的集合。
+
+        参数:
+            messages: 消息字典列表，每个字典至少包含 ``role`` 和 ``content`` 键（兼容 OpenAI 格式）。
+
+        返回:
+            :class:`MultiModalDecision` 值的集合。如果未检测到特殊能力，
+            返回 ``{PLAIN_TEXT}``。
+        """
+        if not messages:
+            return {MultiModalDecision.PLAIN_TEXT}
+
+        decisions: set[MultiModalDecision] = set()
+
+        for msg in messages:
+            role = msg.get("role", "")
+            content = msg.get("content", "")
+
+            # --- VISION: content is a list of blocks with image type ---
+            if isinstance(content, list):
+                for block in content:
+                    if isinstance(block, dict):
+                        block_type = block.get("type", "")
+                        if block_type in ("image_url", "image"):
+                            decisions.add(MultiModalDecision.VISION)
+
+            # --- TOOL_USE: assistant sent tool_calls, or message is tool result ---
+            if msg.get("tool_calls"):
+                decisions.add(MultiModalDecision.TOOL_USE)
+            if role == "tool":
+                decisions.add(MultiModalDecision.TOOL_USE)
+
+            # --- THINKING: assistant included reasoning ---
+            if msg.get("reasoning"):
+                decisions.add(MultiModalDecision.THINKING)
+
+        return decisions if decisions else {MultiModalDecision.PLAIN_TEXT}