From 674519709e10efdd79cca2f5460d34cf91b78452 Mon Sep 17 00:00:00 2001 From: jimi <1847930177@qq.com> Date: Mon, 2 Mar 2026 20:18:51 +0800 Subject: [PATCH] refactor: remove hardcoded routing rules and centralize AI master rules --- qingjian_cs/app/agents.py | 12 ++-- qingjian_cs/app/client.py | 35 ++-------- qingjian_cs/app/orchestrator.py | 51 ++------------- qingjian_cs/app/rules.py | 98 +++++++++++++--------------- qingjian_cs/app/state_machine.py | 15 +---- qingjian_cs/scripts/replay_golden.py | 34 +--------- 6 files changed, 68 insertions(+), 177 deletions(-) diff --git a/qingjian_cs/app/agents.py b/qingjian_cs/app/agents.py index 4d94f0d..bbbee3f 100644 --- a/qingjian_cs/app/agents.py +++ b/qingjian_cs/app/agents.py @@ -115,7 +115,9 @@ class RouterAgent(_AgentRuntime): def __init__(self) -> None: super().__init__( "RouterAgent", - "你是客服路由Agent。只输出路由，不回复客户。必须先调用工具读取意图/风险/订单后再路由。", + rules_prompt() + + "\n你是路由Agent。只输出路由 pre_sales/quote/after_sales/risk，不直接回复客户。" + + " 你必须基于上下文语义路由，禁止关键词硬匹配。", ) async def route(self, context: dict[str, Any]) -> tuple[str, str]: @@ -132,7 +134,7 @@ class QuoteAgent(_AgentRuntime): def __init__(self) -> None: super().__init__( "QuoteAgent", - rules_prompt() + "\n你是报价专家Agent。必须结合图片数量、尺寸和订单状态给出报价动作。", + rules_prompt() + "\n你是报价Agent。负责收图、报价触发、报价回复和报价阶段状态更新。", ) async def decide(self, context: dict[str, Any]) -> Decision: @@ -144,7 +146,7 @@ class AfterSalesAgent(_AgentRuntime): def __init__(self) -> None: super().__init__( "AfterSalesAgent", - rules_prompt() + "\n你是售后专家Agent。优先维护售后状态并给出下一步动作。", + rules_prompt() + "\n你是售后Agent。负责退款/重发/不满意等售后处理与状态推进。", ) async def decide(self, context: dict[str, Any]) -> Decision: @@ -156,7 +158,7 @@ class RiskAgent(_AgentRuntime): def __init__(self) -> None: super().__init__( "RiskAgent", - "你是风控Agent。遇到地图政治/黄暴/外联高风险优先给 transfer 或拒绝性 reply。", + rules_prompt() + "\n你是风控Agent。专注风险识别与风险动作决策。", ) async def decide(self, context: dict[str, Any]) -> Decision: @@ -168,7 +170,7 @@ class PreSalesAgent(_AgentRuntime): def __init__(self) -> None: super().__init__( "PreSalesAgent", - rules_prompt() + "\n你是售前专家Agent。处理打招呼、询价前引导、收图承接。", + rules_prompt() + "\n你是售前Agent。处理咨询承接、收图、澄清需求与转报价前动作。", ) async def decide(self, context: dict[str, Any]) -> Decision: diff --git a/qingjian_cs/app/client.py b/qingjian_cs/app/client.py index d0c28ab..9a74066 100644 --- a/qingjian_cs/app/client.py +++ b/qingjian_cs/app/client.py @@ -16,7 +16,7 @@ from .config import ( from .logger import setup_logger from .observability import activity_event, build_trace_id from .orchestrator import Orchestrator -from .rules import detect_intent, extract_image_urls, prefilter_message +from .rules import extract_image_urls, prefilter_message class QingjianClient: @@ -44,13 +44,8 @@ class QingjianClient: return str(data.get("msg", "") or "").strip() def _debounce_seconds(self, msg: str) -> float: - intent = detect_intent(msg) - if intent == "image": + if extract_image_urls(msg): return 2.5 - if intent in {"pricing", "finish_or_quote_trigger"}: - return 2.0 - if intent == "greeting": - return 1.5 return float(MESSAGE_DEBOUNCE_SECONDS) async def send_message(self, message: dict) -> None: @@ -103,22 +98,6 @@ class QingjianClient: @staticmethod def _humanize_reply(text: str) -> str: t = str(text or "").strip() - replacements = { - "您好呀": "在的", - "您好，": "在的，", - "您好": "在的", - "很高兴为您服务": "我在呢", - "请问您有什么需求": "你要做啥图", - "请问有什么可以帮您": "你要做啥", - "可以把要做的图发我": "图发我就行", - "请您先完成订单付款": "先拍下付款哈", - "麻烦您先完成付款": "先拍下付款哈", - "我们马上为您处理": "我马上处理", - "我这边": "我这", - } - for k, v in replacements.items(): - t = t.replace(k, v) - t = t.replace("~~", "~") return t @staticmethod @@ -133,11 +112,9 @@ class QingjianClient: return False def _fallback_reply(self, action: str) -> str: - if action == "quote": - return "我先看下，马上给你报价。" if action == "transfer": - return "我给你转人工处理哈。" - return "收到，我先看一下哈。" + return "我先给你转人工处理。" + return "收到，我先处理一下。" def _is_outbound_echo(self, data: dict, msg: str) -> bool: """ @@ -177,7 +154,7 @@ class QingjianClient: "goods_name": data.get("goods_name", ""), "goods_order": data.get("goods_order", ""), "msg": merged_msg, - "intent": detect_intent(merged_msg), + "intent": "unknown", "pending_images": len(self.pending_images[key]), "auto_quote_trigger": auto_quote, "last_reply": self.last_reply_key.get(key, ""), @@ -238,7 +215,7 @@ class QingjianClient: try: await asyncio.sleep(AUTO_QUOTE_WAIT_SECONDS) if self.pending_images.get(key): - await self._handle_decision(data, "发完了，报价吧", auto_quote=True) + await self._handle_decision(data, "", auto_quote=True) finally: self.auto_quote_tasks.pop(key, None) diff --git a/qingjian_cs/app/orchestrator.py b/qingjian_cs/app/orchestrator.py index 3122ead..088676d 100644 --- a/qingjian_cs/app/orchestrator.py +++ b/qingjian_cs/app/orchestrator.py @@ -3,15 +3,7 @@ from __future__ import annotations from typing import Any from .agents import AfterSalesAgent, PreSalesAgent, QuoteAgent, RiskAgent, RouterAgent -from .config import FAST_ROUTE_ENABLED from .models import Decision -from .rules import ( - detect_intent, - detect_order_status, - has_map_or_political_risk, - has_porn_risk, - requests_external_contact, -) from .state_machine import evolve_after_sales_state, migrate_state_schema from .store import ConversationStore @@ -31,8 +23,9 @@ class Orchestrator: prev_state = migrate_state_schema(session.get("state", {})) prev_route = session.get("route", "pre_sales") - intent = detect_intent(str(context.get("msg", "") or "")) - order_status = detect_order_status(str(context.get("goods_order", "") or "")) + # 统一改为语义决策：不走关键词意图/订单硬判定。 + intent = "unknown" + order_status = "unknown" merged_ctx = { **context, @@ -42,43 +35,7 @@ class Orchestrator: "order_status": order_status, } - msg = str(context.get("msg", "") or "") - goods_name = str(context.get("goods_name", "") or "") - risk_hit = has_map_or_political_risk(msg, goods_name) or has_porn_risk(msg) or requests_external_contact(msg) - - # 命中硬风控才调用 RiskAgent，避免每条消息都先走一轮模型。 - if risk_hit: - risk_decision = await self.risk.decide(merged_ctx) - route = "risk" - new_state = evolve_after_sales_state( - {**prev_state, **(risk_decision.state_patch or {})}, - route=route, - action=risk_decision.action, - intent=intent, - order_status=order_status, - msg=str(context.get("msg", "") or ""), - ) - self.store.upsert_session(customer_key, context.get("acc_id", ""), context.get("customer_id", ""), route, new_state) - self.store.append_event(customer_key, "decision", {"route": route, "action": risk_decision.action, "reason": risk_decision.reason}) - return route, risk_decision, new_state - - route = "" - route_reason = "" - if FAST_ROUTE_ENABLED: - pending_images = int(context.get("pending_images", 0) or 0) - auto_quote_trigger = bool(context.get("auto_quote_trigger", False)) - if intent in {"pricing", "finish_or_quote_trigger"} and (pending_images > 0 or auto_quote_trigger): - route = "quote" - route_reason = "fast_route_quote_with_pending_images" - elif order_status == "refund": - route = "after_sales" - route_reason = "fast_route_refund" - elif intent in {"image", "greeting", "nonsense", "pricing", "finish_or_quote_trigger", "unknown"}: - route = "pre_sales" - route_reason = "fast_route_common_presales" - - if not route: - route, route_reason = await self.router.route(merged_ctx) + route, route_reason = await self.router.route(merged_ctx) if route == "quote": decision = await self.quote.decide(merged_ctx) diff --git a/qingjian_cs/app/rules.py b/qingjian_cs/app/rules.py index 9bac547..d861353 100644 --- a/qingjian_cs/app/rules.py +++ b/qingjian_cs/app/rules.py @@ -4,15 +4,6 @@ from dataclasses import dataclass IMAGE_URL_RE = re.compile(r"https?://[^\s]+(?:\.jpg|\.jpeg|\.png|\.webp|\.bmp|\.gif)(?:\?[^\s]*)?", re.I) SIZE_RE = re.compile(r"(\d+(?:\.\d+)?)\s*(米|m|M)\s*[xX*乘]\s*(\d+(?:\.\d+)?)\s*(米|m|M)") -MAP_POLITICAL_KWS = ["地图", "国界", "边界", "南海", "台湾", "香港", "澳门", "西藏", "新疆", "政治"] -PORN_RISK_KWS = ["裸", "成人视频", "成人视频", "性爱", "激情", "成人视频"] -EXTERNAL_CONTACT_KWS = ["微信", "vx", "vx", "qq", "手机号", "电话", "加我", "私下"] -PRICE_KWS = ["多少钱", "怎么收费", "报价", "价格", "多少米", "多少"] -GREETING_KWS = ["你好", "您好", "在吗", "在不在", "hello", "hi"] -FINISH_KWS = ["发完了", "没了", "就这些", "报价吧", "可以报价", "先这样"] -NONSENSE_KWS = ["嗯", "哦", "好的", "ok", "1", "收到"] - - @dataclass class RuleResult: ignore: bool = False @@ -30,13 +21,7 @@ def extract_customer_text_from_shop_card(msg: str) -> str: def detect_order_status(order_text: str) -> str: - t = (order_text or "") - if "买家已付款" in t: - return "paid" - if "等待买家付款" in t or "待付款" in t: - return "pending_payment" - if "已退款" in t or "退款" in t: - return "refund" + # 订单状态交给主决策 AI 从上下文语义判断。 return "unknown" @@ -50,25 +35,23 @@ def extract_size_pairs_m(msg: str) -> list[tuple[float, float]]: def has_map_or_political_risk(msg: str, goods_name: str = "") -> bool: - t = f"{msg or ''} {goods_name or ''}".lower() - return any(k.lower() in t for k in MAP_POLITICAL_KWS) + # 风险由 RiskAgent 语义判断。 + return False def has_porn_risk(msg: str) -> bool: - t = (msg or "").lower() - return any(k.lower() in t for k in PORN_RISK_KWS) + # 风险由 RiskAgent 语义判断。 + return False def requests_external_contact(msg: str) -> bool: - t = (msg or "").lower() - return any(k.lower() in t for k in EXTERNAL_CONTACT_KWS) + # 外联风险由 RiskAgent 语义判断。 + return False def is_meaningless_short(msg: str) -> bool: - t = (msg or "").strip().lower() - if len(t) <= 2: - return True - return t in NONSENSE_KWS + # 无意义短句由主决策 AI 语义判断。 + return False def prefilter_message(msg: str, msg_type: int) -> RuleResult: @@ -93,16 +76,7 @@ def detect_intent(msg: str) -> str: m = (msg or "").lower() if IMAGE_URL_RE.search(m): return "image" - if any(k in m for k in FINISH_KWS): - return "finish_or_quote_trigger" - if any(k in m for k in PRICE_KWS): - return "pricing" - if any(k in m for k in GREETING_KWS): - return "greeting" - if requests_external_contact(m): - return "external_contact" - if is_meaningless_short(m): - return "nonsense" + # 其余意图交给 AI 语义判断。 return "unknown" @@ -112,22 +86,40 @@ def extract_image_urls(msg: str) -> list[str]: def rules_prompt() -> str: return ( - "你是淘宝客服主决策。只输出JSON动作，不要解释。\n" - "动作 action 只能是: reply / quote / transfer / noop。\n" - "规则提炼(严格执行):\n" - "1) 客户发图: 先承接, 允许继续收图。\n" - "2) 询价且有图(当前/待处理): 优先 quote。\n" - "3) 无图询价: reply 承接并引导发图。\n" - "4) 客户说发完了/报价吧/没图: 若有待处理图则 quote。\n" - "5) 外部联系方式请求: reply 站内引导, 不给微信QQ手机号。\n" - "6) 地图/政治/黄暴风险: transfer 或拒绝性 reply。\n" - "7) 仅无意义短句(嗯/哦/ok): 给简短自然承接, 不要长回复。\n" - "8) 避免重复同一句; 若上句语义相同则换表达。\n" - "9) 订单已付款: 可回复已安排; 待付款: 提示先付款。\n" - "10) 尺寸明显超大(如>=2m*2m): 提示需补图/重做边缘, 不要直接承诺一模一样。\n" - "11) 店铺差异化: 按 acc_id/persona 口吻回复, 保持真人聊天。\n" - "12) 最终输出只允许一个动作, 不能混合。\n" - "13) reply 必须很短: 1句为主, 不超过20字, 口语化, 不要客服官话和AI腔。\n" + "你是淘宝图像服务客服系统的统一决策AI。必须按以下 MASTER_RULES 执行。\n" + "只输出 JSON 决策，不要解释过程。\n" + "动作 action 只能是: reply / quote / transfer / noop / update_state。\n\n" + "MASTER_RULES:\n" + "A. 统一动作语义\n" + "1) reply: 直接回复客户。\n" + "2) quote: 触发报价或触发看图后报价流程。\n" + "3) transfer: 转人工，必须给 transfer_msg。\n" + "4) noop: 当前不需要回复。\n" + "5) update_state: 仅更新状态，不对外发消息。\n\n" + "B. 售前与报价\n" + "1) 客户发图: 优先承接，可继续收图；不强行一次报完。\n" + "2) 客户询价且已有图(当前图/待处理图/最近图): 优先 action=quote。\n" + "3) 客户无图询价: action=reply，引导其先发图。\n" + "4) 客户说“发完了/就这些/报价吧”: 若有图则 action=quote。\n" + "5) 不能承诺“一模一样原图必找到”，可说先看图评估。\n" + "6) 尺寸很大或要求高还原时，不夸张承诺，先说明可评估后给结论。\n\n" + "C. 订单阶段\n" + "1) 已付款: 可回复“已安排处理/正在处理/完成后发你确认”。\n" + "2) 待付款: 可提示付款，但不与客户争执；必要时先给预览再引导付款。\n" + "3) 退款/售后诉求: 进入售后语境，保持克制，必要时转人工。\n\n" + "D. 风控与合规\n" + "1) 涉政治/地图边界/黄暴/违规内容: 优先 transfer 或拒绝性 reply。\n" + "2) 客户索要微信/QQ/手机号等站外联系方式: 不外呼，站内引导。\n" + "3) 高风险不确定时，不硬答，给保守回复或转人工。\n\n" + "E. 对话质量\n" + "1) 单次只做一个动作，不混合。\n" + "2) 避免重复同一句话；若语义相同，换表达。\n" + "3) reply 必须短: 优先 1 句，口语化，避免AI腔。\n" + "4) 不要输出思考过程，不要输出 tool_use 文本给客户。\n" + "5) 若上下文不足，先澄清 1 个关键问题，不要连续追问。\n\n" + "F. 店铺人格\n" + "1) 按店铺/账号口吻说话，像真人客服，不要机械模板。\n" + "2) 语气友好直接，不啰嗦，不说“作为AI”。\n\n" "输出格式:\n" - '{"action":"reply|quote|transfer|noop","reply":"","transfer_msg":"","quote_mode":"flush_pending|analyze_current_or_recent|collect_only","reason":""}' + '{"action":"reply|quote|transfer|noop|update_state","reply":"","transfer_msg":"","quote_mode":"flush_pending|analyze_current_or_recent|collect_only","state_patch":{},"reason":""}' ) diff --git a/qingjian_cs/app/state_machine.py b/qingjian_cs/app/state_machine.py index 094a331..eb1e62a 100644 --- a/qingjian_cs/app/state_machine.py +++ b/qingjian_cs/app/state_machine.py @@ -38,7 +38,6 @@ def migrate_state_schema(state: dict[str, Any] | None) -> dict[str, Any]: def evolve_after_sales_state(prev_state: dict[str, Any], *, route: str, action: str, intent: str, order_status: str, msg: str) -> dict[str, Any]: s = migrate_state_schema(prev_state) stage = s.get("after_sales_stage", "new") - text = (msg or "").lower() if action == "transfer" or route == "risk": stage = "transferred" @@ -46,22 +45,14 @@ def evolve_after_sales_state(prev_state: dict[str, Any], *, route: str, action: stage = "quoted" s["quote_count"] = int(s.get("quote_count", 0)) + 1 elif route == "after_sales": - if any(k in text for k in ["退款", "退钱", "退货"]): - stage = "refunding" - elif any(k in text for k in ["做完", "完成", "好了", "发我"]): - stage = "waiting_feedback" - elif any(k in text for k in ["补图", "重发", "再发", "原图"]): - stage = "waiting_material" - elif order_status == "paid": + if order_status == "paid": stage = "processing" + elif stage == "new": + stage = "waiting_material" elif route == "pre_sales": if intent == "image": stage = "waiting_material" - # 终态收敛 - if stage == "waiting_feedback" and any(k in text for k in ["没问题", "可以", "确认", "好评"]): - stage = "done" - s["after_sales_stage"] = stage s["last_intent"] = intent s["last_order_status"] = order_status or "unknown" diff --git a/qingjian_cs/scripts/replay_golden.py b/qingjian_cs/scripts/replay_golden.py index 60d9cc1..fcb5a77 100644 --- a/qingjian_cs/scripts/replay_golden.py +++ b/qingjian_cs/scripts/replay_golden.py @@ -11,35 +11,10 @@ if str(ROOT) not in sys.path: sys.path.insert(0, str(ROOT)) from app.orchestrator import Orchestrator -from app.rules import detect_intent, detect_order_status, has_map_or_political_risk, requests_external_contact -from app.state_machine import evolve_after_sales_state, migrate_state_schema CASES_PATH = ROOT / "golden" / "golden_cases.jsonl" -def heuristic_decide(msg: str, goods_order: str) -> tuple[str, str, dict]: - m = (msg or "") - intent = detect_intent(m) - order_status = detect_order_status(goods_order or "") - - if has_map_or_political_risk(m) or requests_external_contact(m): - route, action = "risk", "transfer" - elif "退款" in m: - route, action = "after_sales", "reply" - elif intent == "image": - route, action = "quote", "quote" - elif any(k in m for k in ["多少钱", "报价", "价格", "怎么收费"]): - route, action = "quote", "reply" - elif order_status == "paid": - route, action = "after_sales", "reply" - else: - route, action = "pre_sales", "reply" - - state = migrate_state_schema({}) - state = evolve_after_sales_state(state, route=route, action=action, intent=intent, order_status=order_status, msg=m) - return route, action, state - - async def full_decide(orch: Orchestrator, cid: str, msg: str, goods_order: str) -> tuple[str, str, dict]: context = { "customer_key": f"demo_acc:{cid}", @@ -58,7 +33,7 @@ async def full_decide(orch: Orchestrator, cid: str, msg: str, goods_order: str) async def main() -> int: parser = argparse.ArgumentParser() - parser.add_argument("--mode", choices=["heuristic", "full"], default="heuristic") + parser.add_argument("--mode", choices=["full"], default="full") parser.add_argument("--cases", default=str(CASES_PATH)) args = parser.parse_args() @@ -69,16 +44,13 @@ async def main() -> int: cases = [json.loads(line) for line in p.read_text(encoding="utf-8").splitlines() if line.strip()] - orch = Orchestrator() if args.mode == "full" else None + orch = Orchestrator() passed = 0 failed = 0 for i, c in enumerate(cases, 1): cid = f"case_{i}" - if args.mode == "full": - route, action, state = await full_decide(orch, cid, c.get("msg", ""), c.get("goods_order", "")) - else: - route, action, state = heuristic_decide(c.get("msg", ""), c.get("goods_order", "")) + route, action, state = await full_decide(orch, cid, c.get("msg", ""), c.get("goods_order", "")) ok = route == c.get("expected_route") and action == c.get("expected_action") if ok and c.get("expected_stage"):