perf: switch to single-stage routing and low-latency defaults

qingjian_cs/app/config.py

@@ -12,8 +12,8 @@ OPENAI_API_KEY = os.getenv("OPENAI_API_KEY", "").strip()
 OPENAI_BASE_URL = os.getenv("OPENAI_BASE_URL", "https://ark.cn-beijing.volces.com/api/v3").strip()
 OPENAI_MODEL_NAME = os.getenv("OPENAI_MODEL_NAME", "doubao-seed-2-0-pro-260215").strip()
 
-MESSAGE_DEBOUNCE_SECONDS = float(os.getenv("MESSAGE_DEBOUNCE_SECONDS", "0.8"))
-IMAGE_MESSAGE_DEBOUNCE_SECONDS = float(os.getenv("IMAGE_MESSAGE_DEBOUNCE_SECONDS", "1.2"))
+MESSAGE_DEBOUNCE_SECONDS = float(os.getenv("MESSAGE_DEBOUNCE_SECONDS", "0.2"))
+IMAGE_MESSAGE_DEBOUNCE_SECONDS = float(os.getenv("IMAGE_MESSAGE_DEBOUNCE_SECONDS", "0.4"))
 AUTO_QUOTE_WAIT_SECONDS = int(os.getenv("AUTO_QUOTE_WAIT_SECONDS", "18"))
 AGENT_MAX_ITERS = int(os.getenv("AGENT_MAX_ITERS", "1"))
 FAST_ROUTE_ENABLED = os.getenv("FAST_ROUTE_ENABLED", "1").strip() in {"1", "true", "True", "yes", "on"}
@@ -47,4 +47,4 @@ AUTO_DRAW_TIMEOUT_SECONDS = int(os.getenv("AUTO_DRAW_TIMEOUT_SECONDS", "25"))
 
 # 并发与超时
 MAX_CONCURRENT_TURNS = int(os.getenv("MAX_CONCURRENT_TURNS", "4"))
-DECISION_TIMEOUT_SECONDS = int(os.getenv("DECISION_TIMEOUT_SECONDS", "60"))
+DECISION_TIMEOUT_SECONDS = int(os.getenv("DECISION_TIMEOUT_SECONDS", "8"))

qingjian_cs/app/orchestrator.py

@@ -2,8 +2,7 @@ from __future__ import annotations
 
 from typing import Any
 
-from .agents import AfterSalesAgent, PreSalesAgent, QuoteAgent, RiskAgent, RouterAgent
-from .config import FAST_ROUTE_ENABLED
+from .agents import AfterSalesAgent, PreSalesAgent, QuoteAgent, RiskAgent
 from .image_quote_analyzer import analyze_image_for_quote
 from .models import Decision
 from .state_machine import evolve_after_sales_state, migrate_state_schema
@@ -12,7 +11,6 @@ from .store import ConversationStore
 
 class Orchestrator:
     def __init__(self) -> None:
-        self.router = RouterAgent()
         self.pre_sales = PreSalesAgent()
         self.quote = QuoteAgent()
         self.after_sales = AfterSalesAgent()
@@ -20,9 +18,7 @@ class Orchestrator:
         self.store = ConversationStore()
 
     @staticmethod
-    def _fast_route(context: dict[str, Any]) -> tuple[str, str] | None:
-        if not FAST_ROUTE_ENABLED:
-            return None
+    def _single_stage_route(context: dict[str, Any]) -> tuple[str, str]:
         msg = str(context.get("msg", "") or "")
         lower = msg.lower()
         if bool(context.get("auto_quote_trigger")):
@@ -34,7 +30,7 @@ class Orchestrator:
             return "quote", "fast:pending_image_and_pricing"
         if any(k in lower for k in ("退款", "退钱", "不满意", "重做", "售后", "返工", "重发")):
             return "after_sales", "fast:after_sales_keyword"
-        return None
+        return "pre_sales", "single_stage:default_pre_sales"
 
     async def decide(self, context: dict[str, Any]) -> tuple[str, Decision, dict[str, Any]]:
         customer_key = context["customer_key"]
@@ -54,11 +50,7 @@ class Orchestrator:
             "order_status": order_status,
         }
 
-        fast = self._fast_route(merged_ctx)
-        if fast is not None:
-            route, route_reason = fast
-        else:
-            route, route_reason = await self.router.route(merged_ctx)
+        route, route_reason = self._single_stage_route(merged_ctx)
 
         if route == "quote":
             latest_image_url = str(context.get("latest_image_url", "") or "").strip()