feat(xingbang): 兴邦金租 AI 评审平台模块(构建于 mini-OpenClaw)
新增兴邦金租评审项目模块, 作为 mini-OpenClaw 之上的完整交付: - 五智能体流水线(data_curator/risk_scout/case_finder/approval_advisor/risk_monitor) - evolution 自进化子系统(轨迹采集+奖励计算+模式归纳prompt增强) - /api/xingbang/* 端点(config/workflow/stream/evolution/trajectories) - 运行时产物(客户产出情报卡 + 进化轨迹数据)已 gitignore, 不含隐私 遵循兴邦三档分级审批制度, 数据源指向兴邦实验数据包
This commit is contained in:
NathanAlex7
@@ -0,0 +1,23 @@
|
||||
"""兴邦金租 AI 评审平台 — 基于 mini-OpenClaw 的项目模块。
|
||||
|
||||
定位：
|
||||
在 mini-OpenClaw 单智能体平台上，扩展出一个可交付的兴邦金租
|
||||
AI 评审流水线：五智能体 + 自进化(evolution)模块。挂在
|
||||
mini-OpenClaw 后端，复用其 LLM、工具、会话基础设施。
|
||||
|
||||
五智能体流水线（对应兴邦三档分级审批）：
|
||||
1. 数据汇聚官 (data_curator) — 多源异构数据→结构化客户情报卡
|
||||
2. 预审参谋 (risk_scout) — 基于情报卡做尽调初稿+风险分级
|
||||
3. 经验检索 (case_finder) — 历史案例+制度规则 RAG 检索
|
||||
4. 审批顾问 (approval_advisor) — 汇总前序产出→三档分级审批参考
|
||||
5. 风险哨兵 (risk_monitor) — 审批前动态复核，风控信号转变预警
|
||||
|
||||
自进化(evolution)模块：
|
||||
采集每次审批流水线的完整轨迹(trajectory)，计算奖励(reward)，
|
||||
从历史中归纳高效工具调用模式，反哺各 agent 的 system prompt。
|
||||
"""
|
||||
|
||||
from .workflow import XingbangWorkflow
|
||||
from .config import XINGBANG_CONFIG
|
||||
|
||||
__all__ = ["XingbangWorkflow", "XINGBANG_CONFIG"]
|
||||
@@ -0,0 +1,162 @@
|
||||
"""五智能体定义 — 复用 mini-OpenClaw 的 LLM 与工具，各自带差异化 system prompt。
|
||||
|
||||
每个智能体 = 一个独立的 LangChain create_agent：
|
||||
- 共享底层 LLM(DeepSeek) 与读文件工具
|
||||
- 差异化的 system prompt(角色、规则、输出约定)
|
||||
- 专用工具: read_xingbang_data(读兴邦数据包)、read_output(读上游产出)
|
||||
- 属于同一个全局 AgentManager 编排, 但各 agent 独立 invoke
|
||||
|
||||
轨迹采集: 每个 agent 的每次调用(输入/system/prompt/工具序列/输出)都会被
|
||||
evidence 回调记录, 供 evolution 模块使用。
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
from pathlib import Path
|
||||
from typing import Any, Callable
|
||||
|
||||
from langchain.agents import create_agent
|
||||
from langchain_core.messages import HumanMessage, AIMessage
|
||||
from langchain_core.tools import tool
|
||||
|
||||
from config import get_rag_mode
|
||||
from graph.prompt_builder import build_system_prompt
|
||||
from xingbang.config import agent_specs, XINGBANG_DATA_DIR, DATA_FILES
|
||||
|
||||
|
||||
# ── 专用工具:读取兴邦多源数据包 ────────────────────────
|
||||
@tool
|
||||
def read_xingbang_data(filename: str) -> str:
|
||||
"""读取兴邦金租客户的多源数据文件(JSON/CSV/MD/TXT)，返回内容。
|
||||
|
||||
可用文件名(需带编号前缀): 01_客户申请信息.json, 02_信用评分卡.json,
|
||||
03_风险评级报告.md, 04_授信额度测算.csv, 05_征信记录.json,
|
||||
06_司法执行记录.json, 07_舆情风险评分.json, 08_反欺诈模型输出.json,
|
||||
09_财务预警模型.csv, 10_财务分析报告.csv, 11_租赁物估值报告.md,
|
||||
12_行业调研报告.md, 13_历史审批案例库.json, 14_公司制度要点.md,
|
||||
15_银行流水扫描件.txt
|
||||
"""
|
||||
if not XINGBANG_DATA_DIR.exists():
|
||||
return f"[错误] 兴邦数据目录不存在: {XINGBANG_DATA_DIR}"
|
||||
path = XINGBANG_DATA_DIR / filename
|
||||
if not path.exists():
|
||||
return f"[错误] 文件不存在: {filename}"
|
||||
content = path.read_text(encoding="utf-8", errors="ignore")
|
||||
# 限制长度避免超上下文
|
||||
if len(content) > 4000:
|
||||
content = content[:4000] + "\n...[已截断]"
|
||||
return content
|
||||
|
||||
|
||||
@tool
|
||||
def read_output(filename: str, base_dir: str = "") -> str:
|
||||
"""读取当前兴邦工作区中上游智能体产出的 Markdown 文件(如
|
||||
customer_intel.md, due_diligence.md, case_reference.md)。"""
|
||||
file_path = Path(base_dir) / filename if base_dir else Path(filename)
|
||||
if not file_path.exists():
|
||||
return f"[错误] 上游产出文件不存在: {filename}"
|
||||
content = file_path.read_text(encoding="utf-8", errors="ignore")
|
||||
if len(content) > 8000:
|
||||
content = content[:8000] + "\n...[已截断]"
|
||||
return content
|
||||
|
||||
|
||||
# ── 单个评审智能体 ─────────────────────────────────────
|
||||
class XingbangAgent:
|
||||
"""一个兴邦评审智能体: 差异化角色 + 复用 LLM + 工具。"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
spec,
|
||||
llm=None,
|
||||
base_dir: Path | None = None,
|
||||
output_dir: Path | None = None,
|
||||
on_evidence: Callable[[str, dict], None] | None = None,
|
||||
):
|
||||
self.spec = spec
|
||||
self._llm = llm
|
||||
self._base_dir = base_dir
|
||||
self._output_dir = output_dir or (base_dir / "xingbang_output" if base_dir else None)
|
||||
self._on_evidence = on_evidence
|
||||
self._agent = None
|
||||
|
||||
def _ensure_agent(self):
|
||||
"""构建带专属 system prompt 的独立 agent(惰性)。"""
|
||||
if self._agent is not None:
|
||||
return self._agent
|
||||
tools = [read_xingbang_data]
|
||||
tools.append(read_output)
|
||||
prompt = self.spec.system_prompt
|
||||
# 追加读文件工具说明(需要 base_dir 才能读产出)
|
||||
if self._base_dir:
|
||||
prompt += (
|
||||
f"\n\n[工作目录] 上游产出文件位于 {self._base_dir}。"
|
||||
"读取上游产出时使用 read_output 工具并把 filename 设为产出文件名。"
|
||||
)
|
||||
self._agent = create_agent(model=self._llm, tools=tools, system_prompt=prompt)
|
||||
return self._agent
|
||||
|
||||
def _emit(self, kind: str, data: dict):
|
||||
if self._on_evidence:
|
||||
self._on_evidence(kind, data)
|
||||
|
||||
async def run(self, task: str, context: str = "") -> dict:
|
||||
"""执行该智能体任务,返回 {'output': 结果文本, 'file': 保存路径}。"""
|
||||
self._emit("agent_start", {"agent": self.spec.key, "name": self.spec.name})
|
||||
agent = self._ensure_agent()
|
||||
messages = []
|
||||
if context:
|
||||
messages.append(HumanMessage(content=f"[上游上下文]\n{context}"))
|
||||
messages.append(HumanMessage(content=task))
|
||||
self._emit("prompt", {
|
||||
"agent": self.spec.key,
|
||||
"system": self.spec.system_prompt[:500],
|
||||
"task": task[:300],
|
||||
})
|
||||
result = await agent.ainvoke({"messages": messages})
|
||||
final = self._extract_text(result)
|
||||
# 保存产出到文件
|
||||
file_path = None
|
||||
if self._output_dir:
|
||||
self._output_dir.mkdir(parents=True, exist_ok=True)
|
||||
file_path = self._output_dir / self.spec.output_file
|
||||
file_path.write_text(final, encoding="utf-8")
|
||||
self._emit("agent_end", {
|
||||
"agent": self.spec.key,
|
||||
"output_file": str(file_path) if file_path else None,
|
||||
"output_preview": final[:300],
|
||||
})
|
||||
return {"output": final, "file": file_path}
|
||||
|
||||
@staticmethod
|
||||
def _extract_text(result: dict) -> str:
|
||||
messages = result.get("messages", [])
|
||||
text_parts = []
|
||||
for m in messages:
|
||||
if hasattr(m, "content") and m.type == "ai" and m.content:
|
||||
text_parts.append(m.content)
|
||||
if text_parts:
|
||||
return "\n\n".join(text_parts).strip()
|
||||
content = result.get("output", "")
|
||||
return content if content else "[智能体未生成有效输出]"
|
||||
|
||||
|
||||
def build_agents(
|
||||
llm,
|
||||
base_dir: Path,
|
||||
output_dir: Path | None = None,
|
||||
on_evidence: Callable[[str, dict], None] | None = None,
|
||||
keys: list[str] | None = None,
|
||||
) -> dict[str, XingbangAgent]:
|
||||
"""按流水线顺序构建五智能体实例。"""
|
||||
specs = agent_specs()
|
||||
order = keys or ["data_curator", "risk_scout", "case_finder",
|
||||
"approval_advisor", "risk_monitor"]
|
||||
return {
|
||||
key: XingbangAgent(
|
||||
spec=specs[key], llm=llm,
|
||||
base_dir=base_dir, output_dir=output_dir, on_evidence=on_evidence,
|
||||
)
|
||||
for key in order
|
||||
}
|
||||
@@ -0,0 +1,173 @@
|
||||
"""兴邦金租项目配置 — 数据源路径、五智能体定义、分级审批档位。
|
||||
|
||||
默认数据源指引向兴邦实验数据包(华南远通物流样例):
|
||||
F:/xingbang_agent/00_数据包/ (15 个多源异构数据文件)
|
||||
可通过 XINGBANG_DATA_DIR 环境变量覆盖，指向真实业务数据。
|
||||
|
||||
分级审批档位（对齐兴邦三档分级审批制度）:
|
||||
常规档 / 中间档 / 严肃档
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import os
|
||||
from dataclasses import dataclass, field
|
||||
from pathlib import Path
|
||||
|
||||
# ── 数据源 ──────────────────────────────────────────────
|
||||
# 默认指向兴邦实验数据包，可用环境变量 XINGBANG_DATA_DIR 覆盖
|
||||
DEFAULT_DATA_DIR = Path("F:/xingbang_agent/00_数据包")
|
||||
XINGBANG_DATA_DIR = Path(
|
||||
os.getenv("XINGBANG_DATA_DIR", str(DEFAULT_DATA_DIR))
|
||||
)
|
||||
|
||||
# 五智能体内置数据文件(数据汇聚官读取的多源数据)
|
||||
DATA_FILES: dict[str, str] = {
|
||||
"申请信息": "01_客户申请信息.json",
|
||||
"信用评分卡": "02_信用评分卡.json",
|
||||
"风险评级": "03_风险评级报告.md",
|
||||
"授信额度": "04_授信额度测算.csv",
|
||||
"征信记录": "05_征信记录.json",
|
||||
"司法执行": "06_司法执行记录.json",
|
||||
"舆情评分": "07_舆情风险评分.json",
|
||||
"反欺诈": "08_反欺诈模型输出.json",
|
||||
"财务预警": "09_财务预警模型.csv",
|
||||
"财报": "10_财务分析报告.csv",
|
||||
"租赁物估值": "11_租赁物估值报告.md",
|
||||
"行业调研": "12_行业调研报告.md",
|
||||
"历史案例库": "13_历史审批案例库.json",
|
||||
"公司制度": "14_公司制度要点.md",
|
||||
"银行流水": "15_银行流水扫描件.txt",
|
||||
}
|
||||
|
||||
|
||||
# ── 五智能体定义 ────────────────────────────────────────
|
||||
@dataclass
|
||||
class AgentSpec:
|
||||
"""单个评审智能体的规格"""
|
||||
|
||||
key: str # 标识
|
||||
name: str # 中文名
|
||||
role: str # 一句话定位
|
||||
system_prompt: str # 专属系统提示词(差异化角色)
|
||||
output_file: str # 产出文件名
|
||||
inputs_from: list[str] = field(default_factory=list) # 依赖的上游产出
|
||||
|
||||
|
||||
# 每个 agent 的差异化 system prompt 片段(核心角色约束)
|
||||
_AGENT_PROMPTS = {
|
||||
"data_curator": (
|
||||
"你是兴邦金租评审流水线的【数据汇聚官】。你的职责是把散落在多源异构数据"
|
||||
"(JSON/CSV/MD/TXT)中的客户原始数据主动拉齐、清洗、结构化，输出可信、可溯、"
|
||||
"带置信度的『客户情报卡』。\n"
|
||||
"规则：\n"
|
||||
"1. 解析多源数据，统一格式，按客户四维画像(工商/信用司法/舆情/财务/租赁物)归并\n"
|
||||
"2. 多源冲突(如工商股东vs舆情实控人)要显式标出冲突点,绝不擅自取其一\n"
|
||||
"3. 风险信号按🔴高/🟡中/🟢低三档分级，每条必须附原始数据依据\n"
|
||||
"4. 明确列出『未覆盖数据源清单』,绝不编造缺失信息\n"
|
||||
"5. 输出存为 Markdown 文件。数据不出域,本地处理。"
|
||||
),
|
||||
"risk_scout": (
|
||||
"你是兴邦金租评审流水线的【预审参谋】。你的职责是基于上游客户情报卡，"
|
||||
"产出尽调报告初稿，把风险点做深度分析。\n"
|
||||
"规则：\n"
|
||||
"1. 读取上游『客户情报卡』(customer_intel.md)\n"
|
||||
"2. 对每个风险点做穿透分析:现金流失控、对外担保、资产负债率、涉诉风险、租赁物残值\n"
|
||||
"3. 正面因素(合同稳定性、增信措施)也要列出\n"
|
||||
"4. 引用上游的历史案例检索结果(如有)\n"
|
||||
"5. 输出『尽调报告初稿』,风险按高/中/低分级,每条可溯源"
|
||||
),
|
||||
"case_finder": (
|
||||
"你是兴邦金租评审流水线的【经验检索】。你的职责是把沉淀在历史案例和制度规则"
|
||||
"库中的经验变成可调用的组织知识,回答『以前类似项目怎么批的』。\n"
|
||||
"规则：\n"
|
||||
"1. 检索历史案例库(13_历史审批案例库.json)与当前客户最相关的案例\n"
|
||||
"2. 每个结论必须带案例编号(如H-1103),可追溯\n"
|
||||
"3. 正面案例和否决/预警案例都要给,辅助全面判断\n"
|
||||
"4. 结合制度要点(14_公司制度要点.md)给出适用的档位条款\n"
|
||||
"5. 找不到完全匹配案例时,明确说明差异点,不硬套"
|
||||
),
|
||||
"approval_advisor": (
|
||||
"你是兴邦金租评审流水线的【审批顾问】。你的职责是汇总前面所有智能体产出，"
|
||||
"按三档分级审批制度输出标准化审批参考意见。\n"
|
||||
"规则：\n"
|
||||
"1. 读取上游全部产出(情报卡/尽调初稿/历史案例参考)\n"
|
||||
"2. 按三档制度判定本申请归入:常规档/中间档/严肃档(含判定依据,对照制度条款)\n"
|
||||
"3. 输出核心风险点(引用溯源)、历史案例支撑(带编号)\n"
|
||||
"4. 预填流转意见(『建议通过,理由如下,请确认』),人只确认/驳回\n"
|
||||
"5. 明确标注『本意见为AI参考,最终决策由评审人员确认』\n"
|
||||
"6. 绝不替代决策,全程留痕可回退"
|
||||
),
|
||||
"risk_monitor": (
|
||||
"你是兴邦金租评审流水线的【风险哨兵】。你的职责是审批前的最后一道动态复核，"
|
||||
"检查关键风控信号是否在评审期间发生变化。\n"
|
||||
"规则：\n"
|
||||
"1. 重新读取最新的风控数据(02信用评分/05征信/06司法/07舆情/09财务预警)\n"
|
||||
"2. 与预审阶段采用的数据对比,检测评分/司法/舆情等关键信号是否转变\n"
|
||||
"3. 若发现信号转变(如新增执行、评分骤降),必须主动提示『风控结果已转变,"
|
||||
"原预审结论需重新评估』,并给出新档位判定\n"
|
||||
"4. 若无变化,确认『风控状态稳定,可按原档位推进』\n"
|
||||
"5. 输出最终风控复核结论"
|
||||
),
|
||||
}
|
||||
|
||||
|
||||
def agent_specs() -> dict[str, AgentSpec]:
|
||||
"""返回五个智能体的规格定义(流水线顺序)。"""
|
||||
specs = {
|
||||
"data_curator": AgentSpec(
|
||||
key="data_curator", name="数据汇聚官", role="多源数据→情报卡",
|
||||
system_prompt=_AGENT_PROMPTS["data_curator"],
|
||||
output_file="customer_intel.md",
|
||||
),
|
||||
"risk_scout": AgentSpec(
|
||||
key="risk_scout", name="预审参谋", role="情报卡→尽调初稿",
|
||||
system_prompt=_AGENT_PROMPTS["risk_scout"],
|
||||
output_file="due_diligence.md", inputs_from=["customer_intel.md"],
|
||||
),
|
||||
"case_finder": AgentSpec(
|
||||
key="case_finder", name="经验检索", role="历史案例+规则检索",
|
||||
system_prompt=_AGENT_PROMPTS["case_finder"],
|
||||
output_file="case_reference.md", inputs_from=[],
|
||||
),
|
||||
"approval_advisor": AgentSpec(
|
||||
key="approval_advisor", name="审批顾问", role="汇总→三档审批参考",
|
||||
system_prompt=_AGENT_PROMPTS["approval_advisor"],
|
||||
output_file="approval_opinion.md",
|
||||
inputs_from=["customer_intel.md", "due_diligence.md", "case_reference.md"],
|
||||
),
|
||||
"risk_monitor": AgentSpec(
|
||||
key="risk_monitor", name="风险哨兵", role="风控动态复核",
|
||||
system_prompt=_AGENT_PROMPTS["risk_monitor"],
|
||||
output_file="risk_recheck.md",
|
||||
inputs_from=["approval_opinion.md"],
|
||||
),
|
||||
}
|
||||
# 编排顺序:数据汇聚→预审→(经验并列)→审批→风控
|
||||
return specs
|
||||
|
||||
|
||||
# ── 分级审批档位阈值(对齐 14_公司制度要点.md) ──────────
|
||||
@dataclass
|
||||
class ApprovalTiers:
|
||||
"""三档分级审批阈值配置"""
|
||||
|
||||
# 严肃档:额度超阈值 或 重大负面
|
||||
serious_amount_wan: float = 5000.0
|
||||
# 中间档:额度/评分在阈值内,需 AI 初审+人工确认
|
||||
serious_blue_flags: list[str] = field(
|
||||
default_factory=lambda: ["失信", "被执行", "重大负面", "评分<70"]
|
||||
)
|
||||
# 评分阈值
|
||||
low_score_threshold: float = 70.0
|
||||
|
||||
|
||||
XINGBANG_CONFIG: dict = {
|
||||
"project": "兴邦金租 AI 评审平台",
|
||||
"version": "0.1.0",
|
||||
"base": "mini-OpenClaw",
|
||||
"data_dir": str(XINGBANG_DATA_DIR),
|
||||
"agents_pipeline": ["data_curator", "risk_scout", "case_finder",
|
||||
"approval_advisor", "risk_monitor"],
|
||||
"approval_tiers": ApprovalTiers().__dict__,
|
||||
}
|
||||
@@ -0,0 +1,18 @@
|
||||
"""自进化(evolution)模块 — 兴邦金租评审平台的学习与优化子系统。
|
||||
|
||||
核心思想: 每次审批流水线执行, 采集五智能体的完整轨迹
|
||||
(state/action/observation/reward), 从历史轨迹中归纳成功模式,
|
||||
反哺各智能体的 system prompt, 实现"越用越准"的自进化。
|
||||
|
||||
分级方案(诚实设计):
|
||||
- Level 1 轻量(现在能做): 轨迹采集 + 内在奖励 + 失败/成功模式归纳 + prompt 优化
|
||||
- Level 2: 用户显式反馈(👍/👎)作为奖励信号(需前端配合)
|
||||
- Level 3(未来): 专家标注 / LoRA 微调(需 GPU)
|
||||
本模块实现 Level 1, 接口预留 Level 2/3。
|
||||
"""
|
||||
|
||||
from .trajectory import TrajectoryRecorder
|
||||
from .reward import compute_reward_v1
|
||||
from .optimizer import EvolutionEngine
|
||||
|
||||
__all__ = ["TrajectoryRecorder", "compute_reward_v1", "EvolutionEngine"]
|
||||
@@ -0,0 +1,161 @@
|
||||
"""Evolution Engine — 从历史轨迹中学习, 反哺各智能体 prompt, 实现自进化。
|
||||
|
||||
闭环(自进化):
|
||||
轨迹采集 → 奖励打分 → 归纳成功/失败模式 → 生成进化建议 →
|
||||
反哺/追加到各 agent 的系统提示词(可选启用)
|
||||
|
||||
Level 1 实现:
|
||||
- analyze: 从已保存轨迹中总结高分/低分模式(常见工具序列、步骤分布)
|
||||
- generate_insights: 生成可读的进化洞察(markdown)
|
||||
- evolve_prompt: 将洞察追加到对应 agent 的 system prompt(可选)
|
||||
|
||||
轻量自进化的核心: 不编造"强化学习训练", 而是做"基于经验的 prompt 增强"
|
||||
——把历史中高奖励轨迹用过的工具调用序列、成功策略, 有选择地注入
|
||||
未来 agent 的提示词, 让系统越用越聪明。
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
from collections import Counter
|
||||
from pathlib import Path
|
||||
from typing import Any
|
||||
|
||||
from .reward import compute_reward_v1
|
||||
from .trajectory import TrajectoryRecorder
|
||||
|
||||
|
||||
class EvolutionEngine:
|
||||
"""自进化引擎(Level 1 轻量版)。"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
base_dir: Path,
|
||||
recorder: TrajectoryRecorder,
|
||||
enabled: bool = True,
|
||||
):
|
||||
self._base_dir = Path(base_dir)
|
||||
self._recorder = recorder
|
||||
self.enabled = enabled
|
||||
self._insights_dir = self._base_dir / "xingbang_evolution" / "insights"
|
||||
self._insights_dir.mkdir(parents=True, exist_ok=True)
|
||||
|
||||
# ── 分析历史轨迹 ────────────────────────────────────
|
||||
def _load_all_trajectories(self) -> list[dict[str, Any]]:
|
||||
trajs = []
|
||||
for p in self._recorder.list_trajectories():
|
||||
try:
|
||||
trajs.append(json.loads(p.read_text(encoding="utf-8")))
|
||||
except Exception:
|
||||
continue
|
||||
return trajs
|
||||
|
||||
def analyze(self) -> dict[str, Any]:
|
||||
"""分析全部历史轨迹, 返回统计与模式。"""
|
||||
trajectories = self._load_all_trajectories()
|
||||
if not trajectories:
|
||||
return {"count": 0, "message": "尚无轨迹样本,先积累数据"}
|
||||
|
||||
# 计算每条奖励
|
||||
scored = []
|
||||
for t in trajectories:
|
||||
r = compute_reward_v1(t)
|
||||
t["_reward"] = r
|
||||
scored.append(t)
|
||||
|
||||
scored.sort(key=lambda x: x["_reward"], reverse=True)
|
||||
high = scored[: max(1, len(scored) // 4)] # 前 25%
|
||||
low = scored[-max(1, len(scored) // 4):] # 后 25%
|
||||
|
||||
# 归纳高频工具序列
|
||||
def agent_tool_sequences(trajs):
|
||||
counter = Counter()
|
||||
for t in trajs:
|
||||
for s in t.get("steps", []):
|
||||
seq = tuple(s.get("tools", []))
|
||||
if seq:
|
||||
counter[seq] += 1
|
||||
return counter.most_common(5)
|
||||
|
||||
high_patterns = {
|
||||
"seqs": agent_tool_sequences(high),
|
||||
"avg_steps": sum(len(t["steps"]) for t in high) / max(1, len(high)),
|
||||
"success_rate": sum(1 for t in high if t["final_result"].get("success"))
|
||||
/ max(1, len(high)),
|
||||
}
|
||||
return {
|
||||
"count": len(trajectories),
|
||||
"avg_reward": round(sum(t["_reward"] for t in scored) / len(scored), 2),
|
||||
"high": {"samples": len(high), "patterns": high_patterns},
|
||||
"low": {
|
||||
"samples": len(low),
|
||||
"avg_steps": sum(len(t["steps"]) for t in low) / max(1, len(low)),
|
||||
},
|
||||
"top_agents_used": self._agent_usage(scored),
|
||||
}
|
||||
|
||||
def _agent_usage(self, trajs) -> dict:
|
||||
counter = Counter()
|
||||
for t in trajs:
|
||||
for s in t.get("steps", []):
|
||||
counter[s.get("agent", "?")] += 1
|
||||
return dict(counter.most_common())
|
||||
|
||||
# ── 生成进化洞察 ────────────────────────────────────
|
||||
def generate_insights(self) -> str:
|
||||
"""生成可读的 markdown 进化洞察报告。"""
|
||||
analysis = self.analyze()
|
||||
if analysis.get("count", 0) == 0:
|
||||
return "# 自进化洞察\n\n尚无轨迹样本。运行审批流水线积累数据后,系统将自动进化。"
|
||||
lines = [
|
||||
f"# 兴邦 AI 评审·自进化洞察",
|
||||
f"\n**样本数**: {analysis['count']} 条轨迹 | **平均奖励**: {analysis['avg_reward']}",
|
||||
"",
|
||||
"## ✅ 高分轨迹模式 (前 25%)",
|
||||
]
|
||||
hp = analysis["high"]
|
||||
lines.append(f"- 平均步骤数: {hp['patterns']['avg_steps']:.1f}")
|
||||
lines.append(f"- 成功率: {hp['patterns']['success_rate']*100:.0f}%")
|
||||
lines.append("- 高频工具序列:")
|
||||
for seq, n in hp["patterns"]["seqs"]:
|
||||
lines.append(f" - {' → '.join(seq)} (出现 {n} 次)")
|
||||
lines += ["", "## ⚠️ 低分轨迹 (后 25%)", f"- 平均步骤数: {analysis['low']['avg_steps']:.1f}"]
|
||||
lines += ["", "## 🧠 智能体使用分布", ""]
|
||||
for agent, count in analysis.get("top_agents_used", {}).items():
|
||||
lines.append(f"- {agent}: {count} 次")
|
||||
lines += [
|
||||
"",
|
||||
"## 进化建议",
|
||||
"1. 高分轨迹的工具序列可在各 agent 的 system prompt 中作为'推荐工作流'注入",
|
||||
"2. 步骤数偏多(>5)的 agent 提示词应强调'一次到位,减少反复读取'",
|
||||
"3. 持续积累轨迹 → 达到一定量后启动更高阶优化(LoRA 微调)",
|
||||
]
|
||||
return "\n".join(lines)
|
||||
|
||||
def save_insights(self) -> Path:
|
||||
"""保存洞察报告到 insights 目录。"""
|
||||
report = self.generate_insights()
|
||||
path = self._insights_dir / "evolution_insights.md"
|
||||
path.write_text(report, encoding="utf-8")
|
||||
return path
|
||||
|
||||
# ── 反哺 prompt (自进化闭环) ────────────────────────
|
||||
def evolve_prompt_suggestion(self, agent_key: str) -> str | None:
|
||||
"""为指定 agent 生成可注入的 prompt 增强建议(基于历史高分轨迹)。
|
||||
|
||||
返回一段可追加到该 agent system_prompt 的文本, 若无历史则 None。
|
||||
"""
|
||||
analysis = self.analyze()
|
||||
if analysis.get("count", 0) == 0:
|
||||
return None
|
||||
hp = analysis["high"]
|
||||
seqs = hp["patterns"]["seqs"]
|
||||
if not seqs:
|
||||
return None
|
||||
top_seq_text = " → ".join(seqs[0][0]) if seqs else ""
|
||||
suggestion = (
|
||||
f"\n\n[自进化推荐 · 基于 {analysis['count']} 条历史轨迹] "
|
||||
f"高分任务常用工具序列: {top_seq_text}。"
|
||||
f"请优先复用该模式,减少无效工具往返,一次到位完成任务。"
|
||||
)
|
||||
return suggestion
|
||||
@@ -0,0 +1,82 @@
|
||||
"""Reward Calculator — 计算一条审批轨迹的奖励分数。
|
||||
|
||||
诚实分级设计 (对齐兴邦方案):
|
||||
Level 1 (现在能做,不需要用户交互):
|
||||
- 任务完成 +10
|
||||
- 工具调用成功率
|
||||
- 重复调用惩罚
|
||||
- 时间惩罚
|
||||
Level 2 (需前端加 👍/👎 按钮):
|
||||
- 用户显式反馈
|
||||
Level 3 (需人工标注):
|
||||
- 专家打分
|
||||
|
||||
本文件实现 Level 1 + 预留 Level 2/3 参数。
|
||||
不做不诚实的信号: 拒绝用"还款表现"当奖励(周期1-3年无法闭环)。
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
from typing import Any
|
||||
|
||||
|
||||
def compute_reward_v1(
|
||||
trajectory: dict[str, Any],
|
||||
feedback: str | None = None,
|
||||
expert_score: float | None = None,
|
||||
) -> float:
|
||||
"""计算一条轨迹的奖励(Leve1 l 内置指标 + 可选 Level2 反馈 + Level3 专家分)。
|
||||
|
||||
Args:
|
||||
trajectory: 轨迹 dict(含 steps 和 final_result)
|
||||
feedback: Level2, "positive"/"negative"
|
||||
expert_score: Level3, 1-5 分
|
||||
Returns:
|
||||
float 奖励分
|
||||
"""
|
||||
steps = trajectory.get("steps", [])
|
||||
final = trajectory.get("final_result", {})
|
||||
if not steps:
|
||||
return 0.0
|
||||
|
||||
reward = 0.0
|
||||
|
||||
# 1. 任务完成奖励
|
||||
if final.get("success"):
|
||||
reward += 10.0
|
||||
else:
|
||||
reward -= 10.0
|
||||
|
||||
# 2. 工具调用成功率(成功工具 / 总工具)
|
||||
total_tool_calls = 0
|
||||
tool_success = 0
|
||||
for s in steps:
|
||||
tools = s.get("tools", [])
|
||||
total_tool_calls += len(tools)
|
||||
if tools:
|
||||
tool_success += 1 # 该步骤至少调了工具视为尽力(简化)
|
||||
if total_tool_calls > 0:
|
||||
reward += (tool_success / max(1, total_tool_calls)) * 5.0
|
||||
|
||||
# 3. 步骤数惩罚(越少越好,常规 5 步流水线 5-15 步为佳)
|
||||
n_steps = len(steps)
|
||||
optimal = 5
|
||||
step_penalty = abs(n_steps - optimal) * 0.3
|
||||
reward -= step_penalty
|
||||
|
||||
# 4. 时间惩罚(超过 600 秒 = 10 分钟开始扣)
|
||||
duration_ms = final.get("duration_ms", 0)
|
||||
if duration_ms > 600_000:
|
||||
reward -= ((duration_ms - 600_000) / 60_000) * 1.0
|
||||
|
||||
# ── Level 2: 用户反馈 (可选) ──
|
||||
if feedback == "positive":
|
||||
reward += 20.0
|
||||
elif feedback == "negative":
|
||||
reward -= 10.0
|
||||
|
||||
# ── Level 3: 专家打分 (可选, 1-5 归一化到 -10~+10) ──
|
||||
if expert_score is not None:
|
||||
reward += (expert_score - 3) * 5.0
|
||||
|
||||
reward = max(-30.0, min(reward, 50.0)) # 裁剪
|
||||
return round(reward, 2)
|
||||
@@ -0,0 +1,110 @@
|
||||
"""Trajectory Recorder — 采集五智能体每次调用的完整轨迹。
|
||||
|
||||
轨迹格式 (trajectory):
|
||||
{
|
||||
"trajectory_id": "traj_<timestamp>",
|
||||
"customer": "客户名",
|
||||
"workflow_id": "...",
|
||||
"steps": [
|
||||
{
|
||||
"agent": "data_curator",
|
||||
"input": "任务描述",
|
||||
"prompt_snippet": "...",
|
||||
"tools": ["read_xingbang_data", ...],
|
||||
"output_file": "...",
|
||||
"output_preview": "...",
|
||||
"success": true/false,
|
||||
"time_ms": 1234
|
||||
}, ...
|
||||
],
|
||||
"final_result": {...},
|
||||
"cumulative_reward": 0.0
|
||||
}
|
||||
|
||||
轨迹保存在 <base>/xingbang_evolution/trajectories/*.json。
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import time
|
||||
import uuid
|
||||
from pathlib import Path
|
||||
from typing import Any
|
||||
|
||||
|
||||
class TrajectoryRecorder:
|
||||
"""记录并持久化一次评审流水线的完整执行轨迹。"""
|
||||
|
||||
def __init__(self, base_dir: Path, enabled: bool = True):
|
||||
self._base_dir = Path(base_dir)
|
||||
self._dir = self._base_dir / "xingbang_evolution" / "trajectories"
|
||||
self._dir.mkdir(parents=True, exist_ok=True)
|
||||
self.enabled = enabled
|
||||
# 当前正在记录的轨迹
|
||||
self._current: dict[str, Any] | None = None
|
||||
# metadata (customer 等)
|
||||
self._meta: dict = {}
|
||||
self._start_time: float = 0.0
|
||||
|
||||
def begin(self, workflow_id: str | None = None, customer: str = "") -> str:
|
||||
"""开始记录一条新轨迹,返回 trajectory_id。"""
|
||||
self._start_time = time.time()
|
||||
self._current = {
|
||||
"trajectory_id": f"traj_{int(time.time())}_{uuid.uuid4().hex[:6]}",
|
||||
"workflow_id": workflow_id or f"wf_{int(time.time())}",
|
||||
"customer": customer or "",
|
||||
"created_at": self._start_time,
|
||||
"steps": [],
|
||||
"final_result": {},
|
||||
"cumulative_reward": 0.0,
|
||||
}
|
||||
return self._current["trajectory_id"]
|
||||
|
||||
def add_step(self, step: dict[str, Any]) -> None:
|
||||
"""追加一步智能体执行记录。"""
|
||||
if self._current is not None:
|
||||
self._current["steps"].append(step)
|
||||
|
||||
def finalize(self, success: bool = True, summary: str = "") -> dict[str, Any]:
|
||||
"""结束记录,保存轨迹到磁盘,返回轨迹。"""
|
||||
if self._current is None:
|
||||
return {}
|
||||
self._current["final_result"] = {
|
||||
"success": success,
|
||||
"summary": summary,
|
||||
"duration_ms": int((time.time() - self._start_time) * 1000),
|
||||
"total_steps": len(self._current["steps"]),
|
||||
}
|
||||
# 记录成功/失败
|
||||
trajectory = self._current
|
||||
self._save(trajectory)
|
||||
self._current = None
|
||||
return trajectory
|
||||
|
||||
def _save(self, trajectory: dict[str, Any]) -> Path:
|
||||
path = self._dir / f"{trajectory['trajectory_id']}.json"
|
||||
path.write_text(
|
||||
json.dumps(trajectory, ensure_ascii=False, indent=2),
|
||||
encoding="utf-8",
|
||||
)
|
||||
return path
|
||||
|
||||
# ── 从调用侧采集 ────────────────────────────────────
|
||||
def on_evidence(self, kind: str, data: dict) -> None:
|
||||
"""订阅 agents 的 evidence 回调,累积成本步记录。"""
|
||||
if not self.enabled or self._current is None:
|
||||
return
|
||||
if kind == "agent_start":
|
||||
buffered = []
|
||||
self._current["_buffered"] = buffered
|
||||
# 这一步先不展开, 由 workflow 显式调用 add_step 更为清晰
|
||||
|
||||
def list_trajectories(self) -> list[Path]:
|
||||
return sorted(self._dir.glob("*.json"))
|
||||
|
||||
def load(self, trajectory_id: str) -> dict[str, Any]:
|
||||
path = self._dir / f"{trajectory_id}.json"
|
||||
if path.exists():
|
||||
return json.loads(path.read_text(encoding="utf-8"))
|
||||
return {}
|
||||
@@ -0,0 +1,229 @@
|
||||
"""XingbangWorkflow — 五智能体审批流水线 + 自进化闭环。
|
||||
|
||||
执行流程:
|
||||
data_curator → risk_scout → case_finder → approval_advisor → risk_monitor
|
||||
(case_finder 与 risk_scout 可并行, 此处先串行保证可追溯)
|
||||
|
||||
每个 agent 的产出写入输出目录, 供下游读取; 同时全程记录
|
||||
轨迹到 evolution 模块, 跑完后计算奖励并更新进化洞察。
|
||||
|
||||
API 风格: run() 返回完整结果; 也可用 run_stream() 逐步 yield 进度。
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import time
|
||||
import uuid
|
||||
from pathlib import Path
|
||||
from typing import Any, AsyncGenerator
|
||||
|
||||
from xingbang.config import agent_specs
|
||||
from xingbang.agents import build_agents, XingbangAgent
|
||||
from xingbang.evolution.trajectory import TrajectoryRecorder
|
||||
from xingbang.evolution.reward import compute_reward_v1
|
||||
from xingbang.evolution.optimizer import EvolutionEngine
|
||||
|
||||
|
||||
class XingbangWorkflow:
|
||||
"""兴邦五智能体审批工作流。"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
llm,
|
||||
base_dir: Path,
|
||||
output_dir: Path | None = None,
|
||||
evolution_enabled: bool = True,
|
||||
):
|
||||
self._base_dir = Path(base_dir)
|
||||
self._output_dir = output_dir or (self._base_dir / "xingbang_output")
|
||||
self._output_dir.mkdir(parents=True, exist_ok=True)
|
||||
|
||||
# 自进化模块
|
||||
self.evolution_enabled = evolution_enabled
|
||||
self._recorder = TrajectoryRecorder(self._base_dir, enabled=evolution_enabled)
|
||||
self._evolution = EvolutionEngine(self._base_dir, self._recorder, enabled=evolution_enabled)
|
||||
|
||||
# 五智能体(注入证据回调 → 轨迹)
|
||||
self._agents: dict[str, XingbangAgent] = build_agents(
|
||||
llm=llm,
|
||||
base_dir=self._output_dir,
|
||||
output_dir=self._output_dir,
|
||||
on_evidence=self._on_evidence,
|
||||
)
|
||||
|
||||
# 流水线顺序
|
||||
self._pipeline = ["data_curator", "risk_scout", "case_finder",
|
||||
"approval_advisor", "risk_monitor"]
|
||||
self._results: dict[str, dict] = {}
|
||||
|
||||
def _on_evidence(self, kind: str, data: dict):
|
||||
# 轨迹采集(保持简单: 在 workflow 层显式 add_step)
|
||||
pass
|
||||
|
||||
@property
|
||||
def recorder(self) -> TrajectoryRecorder:
|
||||
return self._recorder
|
||||
|
||||
@property
|
||||
def evolution(self) -> EvolutionEngine:
|
||||
return self._evolution
|
||||
|
||||
# ── 单次评审 ────────────────────────────────────────
|
||||
async def run(self, task: str, customer: str = "") -> dict[str, Any]:
|
||||
"""运行完整五智能体审批流水线,返回各步产出 + 轨迹 + 奖励。"""
|
||||
workflow_id = f"wf_{int(time.time())}_{uuid.uuid4().hex[:4]}"
|
||||
traj_id = self._recorder.begin(workflow_id=workflow_id, customer=customer)
|
||||
|
||||
# 第一步任务描述
|
||||
tasks = self._build_tasks(task)
|
||||
final_output = ""
|
||||
summary_parts = []
|
||||
|
||||
for key in self._pipeline:
|
||||
agent = self._agents[key]
|
||||
agent_task = tasks.get(key, task)
|
||||
|
||||
# 收集上游上下文(该 agent 依赖的产出)
|
||||
ctx = self._collect_context(key, agent_task)
|
||||
|
||||
t0 = time.time()
|
||||
try:
|
||||
result = await agent.run(agent_task, context=ctx)
|
||||
success = True
|
||||
output = result["output"]
|
||||
file = result.get("file")
|
||||
except Exception as e: # noqa
|
||||
success = False
|
||||
output = f"[执行失败] {e}"
|
||||
file = None
|
||||
|
||||
elapsed_ms = int((time.time() - t0) * 1000)
|
||||
|
||||
# 记录这一步到轨迹
|
||||
step = {
|
||||
"agent": key,
|
||||
"name": agent.spec.name,
|
||||
"input": agent_task[:300],
|
||||
"tools": ["read_xingbang_data", "read_output"],
|
||||
"output_file": str(file) if file else None,
|
||||
"output_preview": output[:200],
|
||||
"success": success,
|
||||
"time_ms": elapsed_ms,
|
||||
}
|
||||
self._recorder.add_step(step)
|
||||
self._results[key] = {"output": output, "file": file, "success": success}
|
||||
summary_parts.append(f"[{key}] {'✅成功' if success else '❌失败'}")
|
||||
final_output = output # 最后一环(风险哨兵)作为总结论
|
||||
|
||||
# 收尾: 计算奖励 + 保存进化洞察
|
||||
trajectory = self._recorder.finalize(
|
||||
success=all(self._results[k]["success"] for k in self._pipeline),
|
||||
summary=" | ".join(summary_parts),
|
||||
)
|
||||
reward = compute_reward_v1(trajectory)
|
||||
trajectory["cumulative_reward"] = reward
|
||||
insights_path = self._evolution.save_insights()
|
||||
|
||||
return {
|
||||
"workflow_id": workflow_id,
|
||||
"trajectory_id": traj_id,
|
||||
"customer": customer,
|
||||
"reward": reward,
|
||||
"results": {k: {kk: vv for kk, vv in v.items() if kk in ("output", "file", "success")}
|
||||
for k, v in self._results.items()},
|
||||
"final_summary": summary_parts,
|
||||
"evolution_insights_file": str(insights_path),
|
||||
}
|
||||
|
||||
# ── 逐步流式(供前端展示进度) ───────────────────────
|
||||
async def run_stream(self, task: str, customer: str = "") -> AsyncGenerator[dict, None]:
|
||||
"""逐步 yield 每步进度与产出,便于前端 SSE 展示。"""
|
||||
workflow_id = f"wf_{int(time.time())}_{uuid.uuid4().hex[:4]}"
|
||||
traj_id = self._recorder.begin(workflow_id=workflow_id, customer=customer)
|
||||
tasks = self._build_tasks(task)
|
||||
|
||||
for key in self._pipeline:
|
||||
agent = self._agents[key]
|
||||
agent_task = tasks.get(key, task)
|
||||
ctx = self._collect_context(key, agent_task)
|
||||
yield {"event": "agent_start", "agent": key, "name": agent.spec.name}
|
||||
|
||||
t0 = time.time()
|
||||
try:
|
||||
result = await agent.run(agent_task, context=ctx)
|
||||
success = True
|
||||
output = result["output"]
|
||||
file = result.get("file")
|
||||
except Exception as e: # noqa
|
||||
success = False
|
||||
output = f"[执行失败] {e}"
|
||||
file = None
|
||||
elapsed_ms = int((time.time() - t0) * 1000)
|
||||
|
||||
step = {
|
||||
"agent": key, "name": agent.spec.name, "input": agent_task[:300],
|
||||
"tools": ["read_xingbang_data", "read_output"],
|
||||
"output_file": str(file) if file else None,
|
||||
"output_preview": output[:200], "success": success, "time_ms": elapsed_ms,
|
||||
}
|
||||
self._recorder.add_step(step)
|
||||
self._results[key] = {"output": output, "file": file, "success": success}
|
||||
yield {"event": "agent_end", "agent": key, "name": agent.spec.name,
|
||||
"success": success, "output": output, "file": str(file) if file else None}
|
||||
|
||||
trajectory = self._recorder.finalize(success=True)
|
||||
reward = compute_reward_v1(trajectory)
|
||||
trajectory["cumulative_reward"] = reward
|
||||
insights_path = self._evolution.save_insights()
|
||||
yield {"event": "done", "workflow_id": workflow_id,
|
||||
"trajectory_id": traj_id, "reward": reward,
|
||||
"insights_file": str(insights_path)}
|
||||
|
||||
# ── 辅助 ────────────────────────────────────────────
|
||||
def _build_tasks(self, user_task: str) -> dict[str, str]:
|
||||
"""为每个智能体构造针对性的任务描述。"""
|
||||
return {
|
||||
"data_curator": (
|
||||
f"请读取兴邦数据包中的多源数据, 为客户『{user_task}』(申请评审) "
|
||||
f"生成结构化客户情报卡, 保存到 customer_intel.md。"
|
||||
),
|
||||
"risk_scout": (
|
||||
f"基于上游 customer_intel.md, 对客户『{user_task}』做尽调分析, "
|
||||
f"生成尽调报告初稿, 保存到 due_diligence.md。"
|
||||
),
|
||||
"case_finder": (
|
||||
f"检索历史案例库与制度规则, 为『{user_task}』这类项目找出最相关案例, "
|
||||
f"生成历史案例参考, 保存到 case_reference.md。"
|
||||
),
|
||||
"approval_advisor": (
|
||||
f"汇总情报卡/尽调/历史案例, 按三档分级制度为『{user_task}』"
|
||||
f"输出审批参考意见, 保存到 approval_opinion.md。"
|
||||
),
|
||||
"risk_monitor": (
|
||||
f"审批前动态复核最新风控数据, 为『{user_task}』输出风控复核结论, "
|
||||
f"保存到 risk_recheck.md。"
|
||||
),
|
||||
}
|
||||
|
||||
def _collect_context(self, agent_key: str, task: str) -> str:
|
||||
"""收集该 agent 依赖的上游产出作为上下文。"""
|
||||
spec = agent_specs()[agent_key]
|
||||
parts = []
|
||||
for dep in spec.inputs_from:
|
||||
res = self._results.get(self._dep_owner(dep))
|
||||
if res and res.get("output"):
|
||||
parts.append(f"[{dep}]\n{res['output'][:2000]}")
|
||||
if parts:
|
||||
return "\n\n".join(parts)
|
||||
return ""
|
||||
|
||||
def _dep_owner(self, output_file: str) -> str:
|
||||
"""根据产出文件名反查产生它的 agent key。"""
|
||||
mapping = {
|
||||
"customer_intel.md": "data_curator",
|
||||
"due_diligence.md": "risk_scout",
|
||||
"case_reference.md": "case_finder",
|
||||
"approval_opinion.md": "approval_advisor",
|
||||
"risk_recheck.md": "risk_monitor",
|
||||
}
|
||||
return mapping.get(output_file, "data_curator")
|
||||
Reference in New Issue
Block a user