652 lines
32 KiB
Python
652 lines
32 KiB
Python
import os
|
||
import re
|
||
import hashlib
|
||
import logging
|
||
import time
|
||
import json
|
||
from typing import List, Optional, Any, Dict
|
||
from pydantic_ai import Agent, RunContext
|
||
from pydantic_ai.models.openai import OpenAIChatModel
|
||
from pydantic_ai.providers.openai import OpenAIProvider
|
||
from core.schema import StandardMessage, StandardResponse
|
||
from core.agent_tools import register_agent_tools, TransferSuccessException
|
||
from services.service_wecom_bot import wecom_bot_service
|
||
|
||
logger = logging.getLogger("cs_agent")
|
||
|
||
# 日志详细程度：设置环境变量 AI_LOG_LEVEL=debug 可获得完整日志
|
||
_LOG_FULL_PROMPT = os.getenv("AI_LOG_LEVEL", "").lower() == "debug"
|
||
_LOG_CLIP_LIMIT = int(os.getenv("AI_LOG_CLIP", "2000")) # 日志截断长度
|
||
|
||
from core.skill_manager import skill_manager
|
||
|
||
|
||
_INTERNAL_TOOL_MARKERS = (
|
||
"【历史记录摘要】",
|
||
"【详细记录】",
|
||
"【订单摘要】",
|
||
"【订单详情】",
|
||
)
|
||
_TRANSFER_COMMAND_RE = re.compile(r"^\s*正在为您转接\|\[转移会话\],[^,\r\n]+,[^\r\n]*\s*$")
|
||
|
||
# 历史记录格式检测模式（AI 转述历史时容易泄露）
|
||
_HISTORY_LEAK_PATTERNS = [
|
||
r'\[\d{4}-\d{2}-\d{2}[^\]]*\]\s*(客户|客服)[：:]', # [2026-03-07 12:00:00] 客户：
|
||
r'\[\d{2}:\d{2}:\d{2}\]\s*(客户|客服|我)[：:]', # [12:00:00] 客户：
|
||
r'(根据|查看|查询|翻看)(历史|聊天|对话)(记录|内容)', # 根据历史记录
|
||
r'历史(记录|对话|消息)(显示|表明|中)', # 历史记录显示
|
||
r'之前的(聊天|对话|记录)(中|里|显示)', # 之前的聊天中
|
||
r'共\d+条(历史|对话)?消息', # 共30条历史消息
|
||
r'订单号[:：]\s*\d{10,}', # 订单号:xxxxxxxxxx
|
||
r'(状态|金额|数量)[:：].*(状态|金额|数量)[:：]', # 状态:xxx 金额:xxx 连续出现
|
||
]
|
||
|
||
_FIND_ORIGINAL_INTENT_KEYWORDS = (
|
||
"找图",
|
||
"找原图",
|
||
"原图",
|
||
"素材",
|
||
"大图",
|
||
"源图",
|
||
)
|
||
|
||
_FIND_ORIGINAL_QUESTION_KEYWORDS = (
|
||
"有吗",
|
||
"有没",
|
||
"有没有",
|
||
"能找吗",
|
||
"找得到吗",
|
||
"能不能找到",
|
||
"能找到吗",
|
||
)
|
||
|
||
_REPAIR_INTENT_KEYWORDS = (
|
||
"修复",
|
||
"高清修复",
|
||
"高清",
|
||
"清晰",
|
||
"清楚",
|
||
"变清晰",
|
||
"修清楚",
|
||
"放大清晰",
|
||
)
|
||
|
||
_IMAGE_ALREADY_SENT_HINT_KEYWORDS = (
|
||
"上面不是发了吗",
|
||
"上面不是有吗",
|
||
"我不是发了吗",
|
||
"前面不是发了吗",
|
||
"前面发了",
|
||
"上面发了",
|
||
"我发过了",
|
||
"不是发了吗",
|
||
"都发了",
|
||
"你没看到吗",
|
||
"聊天记录里有",
|
||
"上面有图",
|
||
)
|
||
|
||
_PAYMENT_LINK_REQUEST_KEYWORDS = (
|
||
"付款链接",
|
||
"支付链接",
|
||
"拍单链接",
|
||
"下单链接",
|
||
"付款吧",
|
||
"发我链接",
|
||
"发个链接",
|
||
"发下链接",
|
||
"发一下链接",
|
||
"给我链接",
|
||
)
|
||
|
||
_FILE_HANDOFF_TRANSFER_KEYWORDS = (
|
||
"发送文件了看到了吗",
|
||
"发文件了看到了吗",
|
||
"文件发了看到了吗",
|
||
"文件收到了吗",
|
||
"文件收到没",
|
||
"文件看到了吗",
|
||
"我把文件发过去了",
|
||
"文件发过去了",
|
||
"给你发文件了",
|
||
"源文件发过去了",
|
||
"文件发你了",
|
||
)
|
||
|
||
_DELIVERY_HANDOFF_HINT_KEYWORDS = (
|
||
"发给我吧原图",
|
||
"原图发给我",
|
||
"把原图发给我",
|
||
"发我吧原图",
|
||
"把文件发给我",
|
||
"把成品发给我",
|
||
"做好了发给我",
|
||
"做好了直接发我",
|
||
"做完了发给我",
|
||
"发过来吧",
|
||
)
|
||
|
||
_DESIGNER_SCHEDULE_QUESTION_KEYWORDS = (
|
||
"几点上班",
|
||
"什么时候上班",
|
||
"大概什么时候上班",
|
||
"一般几点上班",
|
||
"明天几点上班",
|
||
"设计师几点上班",
|
||
"设计师什么时候上班",
|
||
"几点在线",
|
||
"什么时候在线",
|
||
"设计师在吗",
|
||
"上班了没",
|
||
"设计师上班了吗",
|
||
)
|
||
|
||
DESIGNER_WORK_START_HOUR = 9
|
||
DESIGNER_WORK_END_HOUR = 12
|
||
|
||
|
||
def _clip(text: str, limit: int = 1200) -> str:
|
||
if text is None:
|
||
return ""
|
||
text = str(text)
|
||
if len(text) <= limit:
|
||
return text
|
||
return f"{text[:limit]}...(截断, 共{len(text)}字)"
|
||
|
||
|
||
async def _notify_brain_fallback(
|
||
msg: StandardMessage,
|
||
error: Exception,
|
||
history_messages: Optional[List[Dict[str, Any]]] = None,
|
||
) -> None:
|
||
history_messages = history_messages or []
|
||
recent_lines: List[str] = []
|
||
for item in history_messages[-3:]:
|
||
role = str(item.get("role") or "").strip() or "unknown"
|
||
content = _clip(str(item.get("content") or "").replace("\r", " ").replace("\n", " "), 80)
|
||
if content:
|
||
recent_lines.append(f"{role}: {content}")
|
||
|
||
current_input = _clip(str(msg.content or "").replace("\r", " ").replace("\n", " "), 200)
|
||
image_count = len(getattr(msg, "image_urls", None) or [])
|
||
lines = [
|
||
"【AI兜底告警】",
|
||
f"店铺：{msg.acc_id or '-'}",
|
||
f"客户：{msg.user_id or '-'}",
|
||
f"消息类型：{getattr(msg, 'msg_type', '-')}",
|
||
f"图片数：{image_count}",
|
||
f"当前消息：{current_input or '-'}",
|
||
f"错误：{_clip(str(error), 300)}",
|
||
]
|
||
if recent_lines:
|
||
lines.append("最近上下文：")
|
||
lines.extend(recent_lines)
|
||
|
||
try:
|
||
ok = await wecom_bot_service.send_text("\n".join(lines))
|
||
if ok:
|
||
logger.info(f"[Brain Fallback Alert] 已发送企业微信告警 user={msg.user_id} acc={msg.acc_id}")
|
||
else:
|
||
logger.warning(f"[Brain Fallback Alert] 企业微信告警发送失败 user={msg.user_id} acc={msg.acc_id}")
|
||
except Exception as notify_err:
|
||
logger.warning(
|
||
f"[Brain Fallback Alert] 企业微信告警异常 user={msg.user_id} acc={msg.acc_id}: {notify_err}"
|
||
)
|
||
|
||
|
||
def _fmt_time(ts: Any) -> str:
|
||
s = str(ts or "").strip()
|
||
if not s:
|
||
return "--:--:--"
|
||
if " " in s:
|
||
return s.split(" ", 1)[1]
|
||
return s
|
||
|
||
|
||
def _sanitize_reply_text(reply_text: str) -> str:
|
||
if not reply_text:
|
||
return ""
|
||
|
||
text = str(reply_text)
|
||
text = re.sub(r'\[\]<\|[^|]+\|>', '', text)
|
||
text = re.sub(r'<\|[^|]*\|>', '', text)
|
||
text = re.sub(r'\[Function[^\]]*\]', '', text)
|
||
text = re.sub(r'\[/?Tool[^\]]*\]', '', text)
|
||
text = re.sub(r'</?tool[_\-]?[^>]*>', '', text, flags=re.IGNORECASE)
|
||
text = re.sub(r'<think[_a-zA-Z0-9]*[^>]*>.*?</think[_a-zA-Z0-9]*[^>]*>', '', text, flags=re.DOTALL)
|
||
text = re.sub(r'<think[_a-zA-Z0-9]*[^>]*>.*', '', text, flags=re.DOTALL)
|
||
text = re.sub(r'</?think[_a-zA-Z0-9]*[^>]*>', '', text)
|
||
text = re.sub(r'```[^`]*```', '', text, flags=re.DOTALL)
|
||
text = re.sub(r'AgentRunResult\([^)]*\)', '', text)
|
||
text = re.sub(r'\[/?[A-Z][a-zA-Z]*(?:Call|End|Start|Result|Return)[^\]]*\]', '', text)
|
||
text = re.sub(r'^\s*\[\s*\{.*$', '', text, flags=re.DOTALL)
|
||
text = re.sub(r'^\s*[\[{].*"name"\s*:.*$', '', text, flags=re.DOTALL)
|
||
text = re.sub(r'[\[\]]{2,}', '', text)
|
||
text = text.strip()
|
||
|
||
if _TRANSFER_COMMAND_RE.fullmatch(text):
|
||
return text
|
||
|
||
if "[转移会话]" in text:
|
||
logger.warning("[Brain] 拦截到混入正文的转接指令，降级为安全兜底回复")
|
||
return "我在帮你看记录，稍等哈"
|
||
|
||
# 检查固定标记
|
||
if any(marker in text for marker in _INTERNAL_TOOL_MARKERS):
|
||
logger.warning("[Brain] 拦截到工具原文泄露，降级为安全兜底回复")
|
||
return "我在帮你看记录，稍等哈"
|
||
|
||
# 检查历史记录泄露模式（AI 转述历史内容）
|
||
for pattern in _HISTORY_LEAK_PATTERNS:
|
||
if re.search(pattern, text):
|
||
logger.warning(f"[Brain] 检测到历史记录泄露模式: {pattern[:30]}...")
|
||
return "我在帮你看记录，稍等哈"
|
||
|
||
return text.strip()
|
||
|
||
|
||
def _normalize_text(text: Any) -> str:
|
||
return str(text or "").strip().lower()
|
||
|
||
|
||
def _infer_image_intent(current_text: str, history: Optional[List[dict]] = None) -> str:
|
||
text = _normalize_text(current_text)
|
||
recent_user_text = "\n".join(
|
||
_normalize_text(h.get("content", ""))
|
||
for h in (history or [])[-6:]
|
||
if h.get("role") == "user"
|
||
)
|
||
combined = f"{recent_user_text}\n{text}"
|
||
|
||
if any(k in combined for k in _REPAIR_INTENT_KEYWORDS):
|
||
return "repair"
|
||
|
||
if any(k in combined for k in _FIND_ORIGINAL_INTENT_KEYWORDS):
|
||
return "find_original"
|
||
|
||
if any(k in text for k in _FIND_ORIGINAL_QUESTION_KEYWORDS):
|
||
return "find_original"
|
||
|
||
return ""
|
||
|
||
|
||
def _history_has_customer_image(history: Optional[List[dict]] = None) -> bool:
|
||
for item in history or []:
|
||
if item.get("role") != "user":
|
||
continue
|
||
msg_type = int(item.get("msg_type") or 0)
|
||
image_urls = item.get("image_urls") or []
|
||
if isinstance(image_urls, str):
|
||
image_urls = [part for part in image_urls.splitlines() if part.strip()]
|
||
content = str(item.get("content") or "")
|
||
if msg_type == 1 or image_urls or ("已收到" in content and "图" in content):
|
||
return True
|
||
return False
|
||
|
||
|
||
def _customer_claims_image_already_sent(current_text: str, history: Optional[List[dict]] = None) -> bool:
|
||
text = _normalize_text(current_text)
|
||
if not text or not _history_has_customer_image(history):
|
||
return False
|
||
return any(keyword in text for keyword in _IMAGE_ALREADY_SENT_HINT_KEYWORDS)
|
||
|
||
|
||
def _requests_payment_link(current_text: str, history: Optional[List[dict]] = None) -> bool:
|
||
text = _normalize_text(current_text)
|
||
if not text:
|
||
return False
|
||
if any(keyword in text for keyword in _PAYMENT_LINK_REQUEST_KEYWORDS):
|
||
return True
|
||
return ("付款" in text or "支付" in text or "拍单" in text or "下单" in text) and "链接" in text
|
||
|
||
|
||
def _history_has_transfer_or_order(history: Optional[List[dict]] = None) -> bool:
|
||
for item in history or []:
|
||
content = str(item.get("content") or "")
|
||
if not content:
|
||
continue
|
||
if "[转移会话]" in content or "设计师上线了" in content:
|
||
return True
|
||
if "[系统订单信息]" in content or "订单状态:" in content or "订单号:" in content:
|
||
return True
|
||
return False
|
||
|
||
|
||
def _requests_file_handoff_transfer(current_text: str, history: Optional[List[dict]] = None) -> bool:
|
||
text = _normalize_text(current_text)
|
||
if not text:
|
||
return False
|
||
if any(keyword in text for keyword in _FILE_HANDOFF_TRANSFER_KEYWORDS):
|
||
return True
|
||
if any(keyword in text for keyword in _DELIVERY_HANDOFF_HINT_KEYWORDS):
|
||
return True
|
||
|
||
has_file_signal = "文件" in text or "源文件" in text
|
||
has_delivery_signal = any(token in text for token in ("发了", "发送了", "发过去", "发你", "给你发"))
|
||
has_receipt_signal = any(token in text for token in ("看到了吗", "收到了吗", "收到没", "看见了吗", "收到吗"))
|
||
if has_file_signal and (has_delivery_signal or has_receipt_signal):
|
||
return True
|
||
|
||
if _history_has_transfer_or_order(history):
|
||
has_asset_signal = any(token in text for token in ("原图", "源文件", "文件", "成品", "图", "图片"))
|
||
has_send_request = any(
|
||
token in text
|
||
for token in ("发给我", "发我", "发过来", "给我吧", "给我发", "直接发", "传给我", "传过来")
|
||
)
|
||
if has_asset_signal and has_send_request:
|
||
return True
|
||
|
||
return False
|
||
|
||
|
||
def _asks_designer_schedule(current_text: str, history: Optional[List[dict]] = None) -> bool:
|
||
text = _normalize_text(current_text)
|
||
if not text:
|
||
return False
|
||
if any(keyword in text for keyword in _DESIGNER_SCHEDULE_QUESTION_KEYWORDS):
|
||
return True
|
||
return ("上班" in text or "在线" in text) and ("几点" in text or "什么时候" in text or "在吗" in text)
|
||
|
||
|
||
class CustomerServiceBrain:
|
||
"""
|
||
重构后的单一 Agent 大脑：
|
||
【全能终极版】统一称呼为“设计师”，支持下线安抚。
|
||
"""
|
||
|
||
def __init__(self, model_name: str = None):
|
||
self.api_key = os.getenv("OPENAI_API_KEY")
|
||
self.base_url = os.getenv("OPENAI_BASE_URL")
|
||
self.model_name = model_name or os.getenv("OPENAI_MODEL", "gpt-4o-mini")
|
||
|
||
model = OpenAIChatModel(
|
||
model_name=self.model_name,
|
||
provider=OpenAIProvider(api_key=self.api_key, base_url=self.base_url)
|
||
)
|
||
|
||
exclude_names = os.getenv("SKILL_EXCLUDE_FROM_PROMPT", "pricing-skill")
|
||
excluded_skills = [s.strip().lower() for s in exclude_names.split(",") if s.strip()]
|
||
all_skills = skill_manager.get_all_skills_text(exclude=excluded_skills)
|
||
logger.info(f"[SkillManager] 已从提示词排除技能: {excluded_skills}")
|
||
|
||
# --- 统一口径后的 System Prompt ---
|
||
system_prompt = (
|
||
"你是一位专注【高清修复】和【找原图】的专业店主。性格干脆，说话自然、专业。\n\n"
|
||
|
||
"【统一称呼规范 - 第一人称原则】\n"
|
||
"1. 你就是店主本人，未转接设计师之前，所有回复必须用第一人称：'我'、'我这边'。\n"
|
||
"2. 例如：客户问进度 → '我在看哈，稍等'；客户催 → '我帮你催下哈'。\n"
|
||
"3. 只有在需要转接时才提'设计师'：'我叫设计师来看下哈'。\n"
|
||
"4. 严禁使用'师傅'、'客服'、'专员'等词汇。\n\n"
|
||
|
||
"【★★★ 历史记录查询 - 最高优先级 ★★★】\n"
|
||
"你有一个 lookup_chat_history_tool 工具，可以查询客户的完整历史聊天记录。\n"
|
||
"以下情况你【必须】先调用此工具查历史，再回复：\n"
|
||
"1. 客户说'之前聊过'、'上次'、'你看聊天记录'、'我发过了'、'前面发了'等\n"
|
||
"2. 客户追问进度：'做好了吗'、'多久能好'、'怎么样了'\n"
|
||
"3. 客户表达不满或困惑：'？'、'你瞎么'、'搞笑'、'说过了'\n"
|
||
"4. 【近期对话回顾】中显示客户之前已发过图或说过需求\n"
|
||
"查到历史后，根据历史内容回复，绝对不要再重复问客户已经回答过的问题！\n\n"
|
||
|
||
"【订单查询工具】\n"
|
||
"你有一个 lookup_customer_orders_tool 工具，可以查询客户的订单记录。\n"
|
||
"以下情况你【必须】调用此工具：\n"
|
||
"1. 客户问'我付款了'、'订单怎么样了'、'发货了吗'\n"
|
||
"2. 客户提到订单号\n"
|
||
"3. 你需要确认客户是否已付款再决定如何回复\n"
|
||
"查到订单后，根据订单状态回复（已付款→'收到，马上安排'；已发货→'已经发了哈'）。\n\n"
|
||
|
||
"【核心逻辑】\n"
|
||
"1. 业务：只聊高清修复和找原图。核心链路：引导发图 -> 问需求 -> 找设计师。\n"
|
||
"2. **主动引导**：只有当客户【从未发过图】且没有历史图片记录时，才引导发图。\n"
|
||
"2.1 **消息延迟安抚**：如果客户说'上面不是发了吗'、'我发过了'、'你没看到吗'，说明他在提醒你图早就发过了。\n"
|
||
" 这时先道歉，类似'不好意思哈，刚刚消息慢了点'，再承接后续；严禁让客户重发图片。\n"
|
||
"3. **非业务问题**：如果客户问招聘、合作、闲聊等与做图无关的话题，礼貌拒绝。\n"
|
||
"4. **客户说没有参考图**：直接转人工：'好的，我这就叫设计师帮您找哈'。\n"
|
||
"5. **客户问尺寸/能否打印/退款**：直接转人工：'这个设计师帮您看下哈'。\n"
|
||
"6. **付款链接特判**：客户明确说'发付款链接'、'支付链接'、'拍单链接'、'下单链接'时，视为强成交信号，必须立即调用转人工工具；严禁只回复'直接下单'。\n"
|
||
"7. **设计师上班时间特判**：客户问'几点上班'、'什么时候上班'、'设计师在吗'、'什么时候在线'时，默认是在问设计师。不要按闲聊处理，也不要回复'我只处理业务'。\n"
|
||
" 设计师固定是早上9点上班，12点下班。应结合当前时间自然回答，不要机械复读。\n"
|
||
"8. **转接时机（严格两步）**：除付款链接特判、文件交接特判外，必须同时满足【有图】+【客户明确或可直接判断的需求】才能转接。\n"
|
||
" 客户只发了图但没说需求 → 先问'亲亲这张是找原图还是修复哈？'\n"
|
||
" 客户说了'有吗'、'能找吗'、'找图'、'找原图'、'有大图吗' → 直接按【找原图】意图处理，不要重复追问。\n"
|
||
" 客户说了'修复'、'高清'、'清晰点'、'放大清晰' → 直接按【高清修复】意图处理，不要重复追问。\n"
|
||
" 客户说'文件发过去了'、'发送文件了看到了吗'、'源文件发你了'、'发给我吧原图'、'做好了直接发我'这类交付话术 → 视为设计师成品/文件交接场景，必须立即调用转人工工具，不要再问客户发图或问需求。\n"
|
||
"9. **下线安抚**：只有工具返回ERROR时才能提设计师不在。根据错误码区分：\n"
|
||
" - ERROR_DESIGNER_NOT_STARTED → 说'还没上班，记下了上班马上处理'（严禁说下班）\n"
|
||
" - ERROR_DESIGNER_OFFLINE → 说'下班了，需求记下明天回'\n"
|
||
" - ERROR_DESIGNER_BUSY → 说'稍等，我帮你联系下'（严禁说下班）\n"
|
||
"10. 正在转接中：如果系统提示已在转接，回：'已经在帮你催了哈，稍等下！'。\n"
|
||
"11. **每次转接必须调用工具**：不要猜测，每次都重新调用。\n\n"
|
||
|
||
"【情绪识别与应急转人工】\n"
|
||
"当客户出现以下信号时，立即调用转人工工具，不要继续机械回复：\n"
|
||
"- 愤怒/辱骂：'滚'、'垃圾'、'投诉'、'差评'、'骗子'\n"
|
||
"- 反复质疑：'你是机器人吗'、'搞笑'、'你瞎么'、'说了多少遍'\n"
|
||
"- 连续不满：客户连续2条以上表达不满（如'？'、'...'、质问语气）\n"
|
||
"转人工话术：'亲亲抱歉，我马上叫设计师亲自来处理哈'\n\n"
|
||
|
||
"【确认短句收尾规则 - 千牛要求最后一句必须是客服说的】\n"
|
||
"客户说'嗯'、'好'、'好的'、'行'、'ok'、'哦'、'知道了'等确认短句时，\n"
|
||
"必须回一句自然的收尾，但严禁复读'嗯咯'！根据上下文选择合适的收尾：\n"
|
||
"- 如果刚谈完需求/报价 → '有问题随时找我哈'\n"
|
||
"- 如果刚说了等设计师 → '好的，有消息马上告诉你'\n"
|
||
"- 如果是闲聊结束 → '好嘞~'\n"
|
||
"每次收尾话术不能重复，要自然变化。\n\n"
|
||
|
||
"【必杀令 - 严格遵守】\n"
|
||
"1. 每句回复严禁超过15个字！语气淘宝亲切风，多用'哈'、'呢'。\n"
|
||
"2. 严禁报价，严禁复读图片已收到的情况。\n"
|
||
"3. 必须原样输出工具返回的'正在为您转接|'指令。\n"
|
||
"4. **严禁**说'在呢铁子'！只能说'在呢'或'在呢亲'。\n"
|
||
"5. **严禁**连续两次回复相同或相似内容！回顾你最近说过的话，换一种说法。\n"
|
||
"6. **严禁**输出任何代码、标记、括号等乱码！只输出自然语言。\n"
|
||
"7. **严禁**自己臆造'下班'！只有工具返回ERROR才能说下班。\n"
|
||
"8. **严禁**在客户已发过图的情况下还说'先发图来看看'！先查历史确认。\n\n"
|
||
|
||
f"业务参考：\n{all_skills}"
|
||
)
|
||
|
||
self.agent = Agent(model=model, system_prompt=system_prompt)
|
||
register_agent_tools(self.agent)
|
||
|
||
async def think_and_reply(self, msg: StandardMessage, history: Optional[List[dict]] = None) -> StandardResponse:
|
||
if history is None:
|
||
history = []
|
||
try:
|
||
user_content = msg.content or ""
|
||
|
||
if _requests_payment_link(user_content, history):
|
||
user_content = (
|
||
"【系统通知：客户正在明确索要付款/支付链接，这是强成交信号。"
|
||
"不要只回复'直接下单'或'平台拍单'，必须立即调用转人工工具转接设计师跟进付款。】\n"
|
||
f"{user_content}"
|
||
)
|
||
elif _requests_file_handoff_transfer(user_content, history):
|
||
logger.info(f"[Brain] 已识别为文件交接转接意图: user={msg.user_id}")
|
||
user_content = (
|
||
"【系统通知：客户现在是在说文件/原图/成品的交接，或者让你把做好的内容直接发过去。"
|
||
"这通常代表设计师已经做完，客户现在是在催交付或确认文件收发。"
|
||
"不要让客户重发图，不要继续问需求，必须立即调用转人工工具转接设计师跟进交付。】\n"
|
||
f"{user_content}"
|
||
)
|
||
elif _asks_designer_schedule(user_content, history):
|
||
now_dt = datetime.now()
|
||
user_content = (
|
||
"【系统通知：客户现在是在问设计师几点上班、什么时候在线、有没有在。"
|
||
"这是有效业务上下文，不要按闲聊或无关业务拒绝。"
|
||
f"设计师固定工作时间是每天{DESIGNER_WORK_START_HOUR}点上班，{DESIGNER_WORK_END_HOUR}点下班。"
|
||
f"当前时间是{now_dt.strftime('%Y-%m-%d %H:%M:%S')}。"
|
||
"请结合当前时间自然回答："
|
||
f"如果现在还没到{DESIGNER_WORK_START_HOUR}点，就表达还没上班，上班后马上处理；"
|
||
f"如果现在已经在{DESIGNER_WORK_START_HOUR}点到{DESIGNER_WORK_END_HOUR}点之间，就表达设计师已经在了或陆续在处理；"
|
||
f"如果现在已经过了{DESIGNER_WORK_END_HOUR}点，就表达已经下班，明天{DESIGNER_WORK_START_HOUR}点后处理。"
|
||
"不要机械照抄这段说明，要自然一点。】\n"
|
||
f"{user_content}"
|
||
)
|
||
|
||
# 客户已发图：告知 AI 图已收到，引导问需求，但不要直接转接
|
||
has_image_message = bool(msg.image_urls) or msg.msg_type == 1
|
||
if not has_image_message and _customer_claims_image_already_sent(user_content, history):
|
||
inferred_intent = _infer_image_intent(user_content, history)
|
||
if inferred_intent == "find_original":
|
||
next_step = "客户当前更像是在问找原图，别再问他有没有发图。"
|
||
elif inferred_intent == "repair":
|
||
next_step = "客户当前更像是在问高清修复，别再问他有没有发图。"
|
||
else:
|
||
next_step = "如果客户需求还不明确，只问这是找原图还是修复，不要让客户重发。"
|
||
user_content = (
|
||
"【系统通知：客户是在提醒你他上面已经发过图片了，可能刚刚网络或消息同步有点慢。"
|
||
"回复时先简短道歉，表示现在已经看到图了，再继续正常承接。"
|
||
f"{next_step}】\n{user_content}"
|
||
)
|
||
if has_image_message:
|
||
image_count = max(len(msg.image_urls), 1)
|
||
if user_content.startswith("【系统：已收到图片消息"):
|
||
user_content = ""
|
||
inferred_intent = _infer_image_intent(user_content, history)
|
||
if inferred_intent == "find_original":
|
||
logger.info(f"[Brain] 已根据客户表述推断为找原图意图: user={msg.user_id}")
|
||
user_content = (
|
||
f"【系统通知：客户已发送 {image_count} 张图片，图已收到不要再让客户发图。"
|
||
f"系统判断客户当前意图是【找原图】；像'有吗'、'能找吗'、'找图'都算找原图意图。"
|
||
f"不要再追问'找原图还是高清修复'，直接按找原图流程继续；如果信息足够就直接转接。】\n{user_content}"
|
||
)
|
||
elif inferred_intent == "repair":
|
||
logger.info(f"[Brain] 已根据客户表述推断为高清修复意图: user={msg.user_id}")
|
||
user_content = (
|
||
f"【系统通知：客户已发送 {image_count} 张图片，图已收到不要再让客户发图。"
|
||
f"系统判断客户当前意图是【高清修复】；像'修复'、'高清'、'清晰点'都算修复意图。"
|
||
f"不要再追问'找原图还是高清修复'，直接按高清修复流程继续；如果信息足够就直接转接。】\n{user_content}"
|
||
)
|
||
else:
|
||
user_content = (
|
||
f"【系统通知：客户已发送 {image_count} 张图片，图已收到不要再让客户发图。"
|
||
f"你现在必须先问客户：这张是找原图还是高清修复？有什么具体要求？"
|
||
f"等客户明确回答后才能转接，严禁跳过问需求直接转接！】\n{user_content}"
|
||
)
|
||
|
||
recent_context = ""
|
||
if history:
|
||
lines = []
|
||
for h in history[-6:]:
|
||
role = "客户" if h.get("role") == "user" else "我"
|
||
content = h.get("content", "")
|
||
lines.append(f"[{_fmt_time(h.get('timestamp'))}] {role}：{content}")
|
||
recent_context = "【近期对话回顾】\n" + "\n".join(lines) + "\n----------------\n"
|
||
|
||
full_input = f"【当前客户ID：{msg.user_id}】\n{recent_context}现在的对话：{user_content}"
|
||
start_time = time.time()
|
||
|
||
# ===== 详细日志：发给 AI 的提示词 =====
|
||
logger.info(f"[AI提示词] user={msg.user_id} acc={msg.acc_id} images={len(msg.image_urls)}\n{full_input}")
|
||
if history:
|
||
history_preview = "\n".join([f" {h.get('role','?')}: {str(h.get('content',''))[:50]}" for h in history[-4:]])
|
||
logger.info(f"[AI历史上下文] 共{len(history)}条:\n{history_preview}")
|
||
|
||
# 尝试运行 AI，捕获转接成功异常以提前终止
|
||
try:
|
||
result = await self.agent.run(full_input, message_history=history)
|
||
except TransferSuccessException as e:
|
||
# 转接工具成功后立即返回，无需等待 AI 继续生成
|
||
elapsed = time.time() - start_time
|
||
logger.info(f"[Brain] 转接成功（提前终止，耗时{elapsed:.1f}s）: {e.transfer_cmd[:60]}")
|
||
return StandardResponse(
|
||
reply_content=e.transfer_cmd,
|
||
need_transfer=True,
|
||
metadata={"acc_id": msg.acc_id, "acc_type": msg.acc_type}
|
||
)
|
||
|
||
elapsed = time.time() - start_time
|
||
logger.info(f"[Brain] AI处理完成，总耗时{elapsed:.1f}s")
|
||
|
||
# ===== 详细日志：AI 的思考过程和工具调用 =====
|
||
pending_transfer_reason = ""
|
||
pending_transfer_error = ""
|
||
try:
|
||
all_msgs = result.all_messages()
|
||
for idx, m in enumerate(all_msgs):
|
||
msg_kind = getattr(m, 'kind', type(m).__name__)
|
||
if hasattr(m, 'parts'):
|
||
for part in m.parts:
|
||
part_kind = getattr(part, 'part_kind', '')
|
||
if part_kind == 'tool-call':
|
||
tool_name = getattr(part, 'tool_name', '?')
|
||
tool_args = getattr(part, 'args', {})
|
||
logger.info(f"[AI思考] 步骤{idx+1} 调用工具: {tool_name}({tool_args})")
|
||
if tool_name == "transfer_to_human_tool":
|
||
if isinstance(tool_args, str):
|
||
try:
|
||
tool_args = json.loads(tool_args)
|
||
except Exception:
|
||
tool_args = {"reason": tool_args}
|
||
if isinstance(tool_args, dict):
|
||
pending_transfer_reason = str(tool_args.get("reason") or "").strip()
|
||
elif part_kind == 'tool-return':
|
||
content = str(getattr(part, 'content', ''))[:200]
|
||
logger.info(f"[AI思考] 步骤{idx+1} 工具返回: {content}")
|
||
full_content = str(getattr(part, 'content', ''))
|
||
if full_content.startswith("ERROR_DESIGNER_"):
|
||
pending_transfer_error = full_content
|
||
elif part_kind == 'text':
|
||
content = str(getattr(part, 'content', ''))[:150]
|
||
if content.strip():
|
||
logger.info(f"[AI思考] 步骤{idx+1} 文本输出: {content}")
|
||
except Exception as log_err:
|
||
logger.debug(f"[AI思考日志] 解析失败: {log_err}")
|
||
|
||
# --- 转接指令：直接从工具返回截获，不经过 AI 二次加工 ---
|
||
transfer_cmd = ""
|
||
for m in result.all_messages():
|
||
if hasattr(m, 'parts'):
|
||
for part in m.parts:
|
||
if getattr(part, 'part_kind', '') == 'tool-return':
|
||
content = str(getattr(part, 'content', ''))
|
||
if "[转移会话]" in content:
|
||
transfer_cmd = content
|
||
|
||
if transfer_cmd:
|
||
logger.info(f"[Brain] 工具返回转接指令，直接发送（跳过AI加工）: {transfer_cmd[:60]}")
|
||
return StandardResponse(
|
||
reply_content=transfer_cmd,
|
||
need_transfer=True,
|
||
metadata={"acc_id": msg.acc_id, "acc_type": msg.acc_type}
|
||
)
|
||
|
||
# --- 非转接场景：取 AI 的正常回复 ---
|
||
reply_text = ""
|
||
raw_output = getattr(result, 'output', None) or getattr(result, 'data', None)
|
||
if isinstance(raw_output, str):
|
||
reply_text = raw_output
|
||
|
||
# 清理模型泄露的内部标记/工具原文
|
||
reply_text = _sanitize_reply_text(reply_text)
|
||
|
||
# 过滤"在呢铁子"
|
||
if "在呢铁子" in reply_text:
|
||
reply_text = reply_text.replace("在呢铁子", "在呢亲")
|
||
|
||
if not reply_text:
|
||
reply_text = "稍等我看看。"
|
||
|
||
logger.info(f"[THINK/RAW_OUTPUT] user={msg.user_id}\n{_clip(reply_text)}")
|
||
|
||
need_transfer = "[转移会话]" in reply_text
|
||
|
||
return StandardResponse(
|
||
reply_content=reply_text,
|
||
need_transfer=need_transfer,
|
||
metadata={
|
||
"acc_id": msg.acc_id,
|
||
"acc_type": msg.acc_type,
|
||
"pending_transfer": bool(pending_transfer_error and pending_transfer_reason),
|
||
"pending_transfer_reason": pending_transfer_reason,
|
||
"pending_transfer_error": pending_transfer_error,
|
||
}
|
||
)
|
||
|
||
except Exception as e:
|
||
logger.error(f"[Brain Error]: {e}")
|
||
await _notify_brain_fallback(msg, e, history)
|
||
return StandardResponse(reply_content="好哒，我在看图，稍等回你哈。", metadata={"acc_id": msg.acc_id})
|