Enhance workflow routing and agent configuration management

- Introduce WorkflowLLMRouter for pre-response LLM routing, allowing agents to determine the appropriate function to call based on user input. - Implement UserTurnRoutingProcessor to manage user turns before reaching the LLM, ensuring proper routing and handling of user messages. - Refactor WorkflowBrain to integrate new routing logic and enhance agent stage configuration, including entry modes and resource management. - Update service factory to support dynamic LLM resource configuration based on workflow settings. - Add tests for new routing functionality and ensure proper handling of user messages in various scenarios.
2026-07-14 09:36:28 +08:00
parent 32aef14ddb
commit 72856bf3a7
19 changed files with 2611 additions and 552 deletions
--- a/backend/services/workflow_router.py
+++ b/backend/services/workflow_router.py
@@ -0,0 +1,129 @@
+"""Pre-response LLM routing for Workflow Agent edges.
+
+The router deliberately uses a separate, short completion.  Its only output is
+a required function choice, so the current Agent cannot speak before the graph
+has decided whether the user turn belongs to another node.
+"""
+
+from __future__ import annotations
+
+import json
+from collections.abc import Callable
+from typing import Any
+
+from loguru import logger
+from models import AssistantConfig
+from openai import AsyncOpenAI
+
+
+STAY_ON_CURRENT_AGENT = "workflow_stay_on_current_agent"
+MAX_ROUTING_HISTORY_ENTRIES = 20
+
+
+class WorkflowLLMRouter:
+    """Select one LLM edge before the conversational LLM is allowed to reply."""
+
+    def __init__(self, cfg: AssistantConfig):
+        self._cfg = cfg
+
+    async def select_edge(
+        self,
+        *,
+        node_name: str,
+        node_prompt: str,
+        edges: list[dict[str, Any]],
+        history: list[dict[str, str]],
+        variables: dict[str, Any],
+        edge_name: Callable[[dict[str, Any]], str],
+        edge_description: Callable[[dict[str, Any]], str],
+    ) -> str | None:
+        """Return an edge function name, STAY, or None when routing failed."""
+        if not edges:
+            return STAY_ON_CURRENT_AGENT
+
+        names = {edge_name(edge) for edge in edges}
+        stay_name = STAY_ON_CURRENT_AGENT
+        while stay_name in names:
+            stay_name = f"_{stay_name}"
+
+        tools = [
+            {
+                "type": "function",
+                "function": {
+                    "name": edge_name(edge),
+                    "description": edge_description(edge),
+                    "parameters": {"type": "object", "properties": {}},
+                },
+            }
+            for edge in edges
+        ]
+        tools.append(
+            {
+                "type": "function",
+                "function": {
+                    "name": stay_name,
+                    "description": "所有转移条件都不满足，继续由当前 Agent 处理用户消息。",
+                    "parameters": {"type": "object", "properties": {}},
+                },
+            }
+        )
+
+        ordered_conditions = "\n".join(
+            f"{index + 1}. {edge_description(edge)}"
+            for index, edge in enumerate(edges)
+        )
+        router_prompt = (
+            "你是工作流路由器，不是对话助手。收到一轮完整用户输入后，"
+            "必须且只能调用一个提供的函数，禁止输出任何口头回复。\n"
+            "按给出的顺序判断转移条件；选择第一个明确满足的转移函数。"
+            "如果没有条件满足，调用留在当前 Agent 的函数。\n\n"
+            f"当前节点：{node_name}\n"
+            f"当前节点任务：{node_prompt or '未配置'}\n"
+            f"转移条件：\n{ordered_conditions}"
+        )
+        recent_history = history[-MAX_ROUTING_HISTORY_ENTRIES:]
+        routing_input = json.dumps(
+            {
+                "conversation": recent_history,
+                "session_variables": variables,
+            },
+            ensure_ascii=False,
+            separators=(",", ":"),
+        )
+        extra_body = self._cfg.llm_values.get("extraBody")
+        request_extra = (
+            {"extra_body": extra_body} if isinstance(extra_body, dict) else {}
+        )
+        client = AsyncOpenAI(
+            api_key=self._cfg.llm_api_key,
+            base_url=self._cfg.llm_base_url,
+            timeout=15.0,
+        )
+        try:
+            response = await client.chat.completions.create(
+                model=self._cfg.model,
+                messages=[
+                    {"role": "system", "content": router_prompt},
+                    {"role": "user", "content": routing_input},
+                ],
+                tools=tools,
+                tool_choice="required",
+                temperature=0,
+                **request_extra,
+            )
+            tool_calls = response.choices[0].message.tool_calls or []
+            if not tool_calls:
+                logger.warning("Workflow 路由 LLM 未返回函数调用，留在当前 Agent")
+                return STAY_ON_CURRENT_AGENT
+            selected = str(tool_calls[0].function.name or "")
+            if selected == stay_name:
+                return STAY_ON_CURRENT_AGENT
+            if selected not in names:
+                logger.warning(f"Workflow 路由 LLM 返回未知函数:{selected}")
+                return STAY_ON_CURRENT_AGENT
+            return selected
+        except Exception as exc:  # noqa: BLE001 - routing failure must not end the call
+            logger.warning(f"Workflow LLM 边判断失败，留在当前 Agent:{exc}")
+            return None
+        finally:
+            await client.close()