SaaS-AI Agent 节点设计说明书 v1.0 正式版

文档信息

---

1. 项目背景与目标

1.1 核心使命

构建以SaaS为主系统、AI Agent为辅节点的B2B再生塑料CRM平台，实现：

1.2 铁律约束

---

2. 业务问题回答与决策记录

问题1：项目商机分析Agent的数据来源 ✅

**决策**:

问题2：AI Agent模型选择与部署位置 ✅

**决策**:

- **DeepSeek-R1 32B/70B** (推荐首选)：推理能力强，适合商机分析

- **Qwen2.5-72B** (通义千问)：中文理解优秀，适合日报周报汇总

- **ChatGLM3-6B/12B** (智谱)：轻量级，适合实时响应场景

- **Baichuan2-13B** (百川)：中文场景优化

- 主流程API：毫秒级响应

- Agent分析：异步队列，60秒内完成，结果通过企业微信推送

问题3："人类确认"实现方式 ✅

**决策**:

- 非流程性/不影响其他节点：等待直至批复

- 影响其他流程节点：维持旧规则执行，新规则审批通过后切换

问题4：项目商机分析评分标准 ✅

**决策**:

- 客户关系: 30%

- 技术匹配: 25%

- 价格竞争力: 20%

- 切换意愿: 15%

- 时间窗口: 10%

- 项目阶段: 30%

- 客户订单状态: 25%

- 技术匹配: 20%

- 时间规划: 15%

- 竞争态势: 10%

- 高商机: ≥70分

- 中商机: 50-69分

- 低商机: 30-49分

- 放弃: <30分

问题5：SaaS多租户与数据隔离 ✅

**决策**:

---

3. SaaS核心业务流程梳理

3.1 业务流程全景图

┌─────────────────────────────────────────────────────────────┐
│                    TopcentralMall CRM SaaS (多租户)          │
├─────────────────────────────────────────────────────────────┤
│  租户隔离层 (tenant_id 物理隔离)                           │
├─────────────────────────────────────────────────────────────┤
│  客户管理 → 销售过程 → 办公自动化 → 营销分析 → 产品库存    │
│      ↓         ↓          ↓           ↓          ↓          │
│   [Agent]    [Agent]    [Agent]     [Agent]   [Agent]      │
│   客户画像   商机评分   日报汇总    流失预警   库存建议      │
│      ↓         ↓          ↓           ↓          ↓          │
│  企业微信    企业微信   企业微信    企业微信   自动执行      │
│  人类确认    人类确认   人类确认    人类确认                │
│      ↓         ↓          ↓           ↓                    │
│   SaaS执行   SaaS执行  SaaS执行   SaaS执行                │
└─────────────────────────────────────────────────────────────┘

3.2 多租户数据隔离架构

┌─────────────────────────────────────────┐
│           API Gateway (Nginx)           │
│         JWT Token + tenant_id           │
└─────────────────────────────────────────┘
                    │
┌─────────────────────────────────────────┐
│         SaaS Core (FastAPI)             │
│    毫秒级响应，不阻塞主流程              │
└─────────────────────────────────────────┘
                    │
        ┌──────────┼──────────┐
        │          │          │
   ┌────┴────┐┌───┴────┐┌────┴────┐
   │ Tenant  ││ Tenant ││ Tenant  │
   │   A     ││   B    ││   C     │
   │(独立DB) ││(独立DB)││(独立DB) │
   └────┬────┘└───┬────┘└────┬────┘
        │          │          │
   ┌────┴────────────────────┴────┐
   │      Agent Analysis Queue      │
   │   (异步执行，Redis/RabbitMQ)    │
   │   DeepSeek/Qwen/私有模型       │
   └────────────────────────────────┘
                    │
        ┌──────────┴──────────┐
        │                     │
   ┌────┴────┐          ┌────┴────┐
   │ 企业微信 │          │  邮件   │
   │ 推送确认 │          │  通知   │
   └─────────┘          └─────────┘

---

4. Agent节点详细设计

4.1 节点清单 (15个)

4.2 节点A03：项目商机分析Agent (核心节点)

输入上下文 Schema

{
  "tenant_id": "string (required)",
  "workflow_instance_id": "string (required)",
  "node_id": "A03",
  "customer_relation": {
    "relationship_depth": "number: 1-10",
    "decision_maker_relation": "enum: [excellent, good, poor]",
    "tech_relation": "enum: [excellent, good, poor]",
    "cooperation_years": "number",
    "last_visit_date": "date",
    "visit_count_30d": "number"
  },
  "mass_production_project": {
    "current_supplier": "string",
    "switch_willingness": "enum: [strong, moderate, weak, none]",
    "plastic_requirements": "string",
    "mold_structure": "enum: [simple, medium, complex]",
    "molding_process": "string",
    "cert_requirements": "string",
    "monthly_volume": "number",
    "price_range": "string",
    "payment_terms": "enum: [30, 60, 90, 120]",
    "competitors": "string"
  },
  "new_project": {
    "project_stage": "enum: [rfq, sample, mold, ppap, mass]",
    "customer_has_order": "enum: [yes, bidding, no]",
    "sample_date": "date",
    "mold_date": "date",
    "mass_production_date": "date",
    "project_lifecycle": "enum: [short, medium, long]"
  },
  "switch_motivation": {
    "motivation_factors": "array: [price, quality, supply, service, tech, relation]",
    "risk_assessment": "string"
  },
  "historical_data": {
    "quotes": "array",
    "samples": "array",
    "contracts": "array",
    "orders": "array",
    "travel_reports": "array"
  }
}

输出JSON Schema

{
  "tenant_id": "string",
  "workflow_instance_id": "string",
  "node_id": "A03",
  "timestamp": "ISO8601",
  "overall_score": "number: 0-100",
  "score_level": "enum: [high, medium, low, abandon]",
  "score_breakdown": {
    "relation_score": "number: 0-30",
    "tech_match_score": "number: 0-25",
    "price_score": "number: 0-20",
    "switch_willingness_score": "number: 0-15",
    "time_window_score": "number: 0-10"
  },
  "relation_assessment": {
    "score": "number: 0-10",
    "level": "enum: [excellent, good, poor]",
    "risks": ["string"],
    "opportunities": ["string"]
  },
  "mass_project_analysis": {
    "feasibility_score": "number: 0-100",
    "switch_probability": "number: 0-1",
    "competitor_analysis": {
      "primary_competitor": "string",
      "competitor_strengths": ["string"],
      "competitor_weaknesses": ["string"],
      "our_advantages": ["string"]
    },
    "risk_matrix": [
      {
        "risk_type": "string",
        "level": "enum: [high, medium, low]",
        "description": "string"
      }
    ]
  },
  "new_project_analysis": {
    "potential_score": "number: 0-100",
    "timeline_assessment": "string",
    "risks": ["string"],
    "opportunities": ["string"]
  },
  "switch_willingness_analysis": {
    "willingness_level": "enum: [strong, moderate, weak, none]",
    "motivation_analysis": "string",
    "risk_factors": ["string"]
  },
  "strategy_recommendations": [
    {
      "priority": "number: 1-5",
      "title": "string",
      "description": "string",
      "expected_outcome": "string",
      "timeline": "string"
    }
  ],
  "risk_warnings": [
    {
      "severity": "enum: [high, medium, low]",
      "category": "string",
      "description": "string",
      "mitigation": "string"
    }
  ],
  "next_actions": [
    {
      "action": "string",
      "owner": "string",
      "deadline": "date",
      "priority": "enum: [high, medium, low]"
    }
  ],
  "confidence": "number: 0-1",
  "model_version": "string",
  "suggested_human_review": "boolean: true"
}

不可逾越边界

4.3 新建数据库表设计

表1: customer_decision_tree (客户决策树)

CREATE TABLE customer_decision_tree (
    id SERIAL PRIMARY KEY,
    tenant_id VARCHAR(64) NOT NULL,
    customer_id VARCHAR(64) NOT NULL,
    decision_maker_name VARCHAR(128),
    decision_maker_title VARCHAR(64),
    decision_maker_contact VARCHAR(64),
    influence_level INTEGER CHECK (influence_level BETWEEN 1 AND 10),
    relationship_quality VARCHAR(16) CHECK (relationship_quality IN ('excellent', 'good', 'poor')),
    tech_contact_name VARCHAR(128),
    tech_contact_title VARCHAR(64),
    tech_influence_level INTEGER CHECK (tech_influence_level BETWEEN 1 AND 10),
    procurement_contact_name VARCHAR(128),
    procurement_contact_title VARCHAR(64),
    procurement_influence_level INTEGER CHECK (procurement_influence_level BETWEEN 1 AND 10),
    created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
    updated_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
    UNIQUE(tenant_id, customer_id),
    FOREIGN KEY (tenant_id, customer_id) REFERENCES customers(tenant_id, id)
);

表2: project_analysis_results (项目分析结果)

CREATE TABLE project_analysis_results (
    id SERIAL PRIMARY KEY,
    tenant_id VARCHAR(64) NOT NULL,
    workflow_instance_id VARCHAR(128) NOT NULL,
    node_id VARCHAR(16) NOT NULL,
    project_id VARCHAR(64) NOT NULL,
    overall_score INTEGER CHECK (overall_score BETWEEN 0 AND 100),
    score_level VARCHAR(16) CHECK (score_level IN ('high', 'medium', 'low', 'abandon')),
    score_breakdown JSONB NOT NULL,
    relation_assessment JSONB,
    mass_project_analysis JSONB,
    new_project_analysis JSONB,
    switch_willingness_analysis JSONB,
    strategy_recommendations JSONB,
    risk_warnings JSONB,
    next_actions JSONB,
    confidence DECIMAL(3,2),
    model_version VARCHAR(32),
    human_review_status VARCHAR(16) DEFAULT 'pending' CHECK (human_review_status IN ('pending', 'approved', 'rejected', 'modified')),
    human_review_comment TEXT,
    human_reviewer_id VARCHAR(64),
    created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
    FOREIGN KEY (tenant_id, project_id) REFERENCES projects(tenant_id, id)
);

表3: agent_execution_logs (Agent执行日志)

CREATE TABLE agent_execution_logs (
    id SERIAL PRIMARY KEY,
    tenant_id VARCHAR(64) NOT NULL,
    workflow_instance_id VARCHAR(128) NOT NULL,
    node_id VARCHAR(16) NOT NULL,
    model_name VARCHAR(64) NOT NULL,
    model_version VARCHAR(32) NOT NULL,
    input_tokens INTEGER,
    output_tokens INTEGER,
    execution_time_ms INTEGER,
    status VARCHAR(16) CHECK (status IN ('success', 'failure', 'timeout')),
    error_message TEXT,
    created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP
);

---

5. 模型部署架构

5.1 部署方案

┌─────────────────────────────────────────────────────────────┐
│                     阿里云服务器 47.96.253.194              │
├─────────────────────────────────────────────────────────────┤
│  ┌─────────────────┐  ┌─────────────────┐  ┌─────────────┐│
│  │   SaaS Core     │  │   Agent Queue   │  │  Model API  ││
│  │   (FastAPI)     │  │   (Redis/Celery)│  │  (vLLM)     ││
│  │   毫秒级响应    │  │   异步处理      │  │  模型推理   ││
│  └─────────────────┘  └─────────────────┘  └─────────────┘│
│         │                    │                    │        │
│         └────────────────────┴────────────────────┘        │
│                              │                             │
│  ┌─────────────────────────────────────────────────────────┐│
│  │              Nginx (API Gateway)                        ││
│  │         /api/* → SaaS Core                              ││
│  │         /agent/* → Agent Queue                          ││
│  │         /model/* → Model API (内部)                   ││
│  └─────────────────────────────────────────────────────────┘│
└─────────────────────────────────────────────────────────────┘
                              │
                    ┌─────────┴─────────┐
                    │                   │
              ┌─────┴─────┐      ┌─────┴─────┐
              │ DeepSeek  │      │   Qwen    │
              │  R1-32B   │      │ 2.5-72B   │
              │ (本地GPU) │      │ (API调用) │
              └───────────┘      └───────────┘

5.2 模型选择策略

5.3 私有模型定义

**私有模型**指部署在企业自有服务器或私有云环境中的AI模型，特点：

**推荐私有模型**:

---

6. 企业微信确认流程

6.1 确认流程图

Agent生成建议
    │
    ▼
┌─────────────────┐
│ 企业微信推送    │
│ • 建议摘要      │
│ • 关键数据      │
│ • 确认按钮      │
│ • 驳回按钮      │
│ • 查看详情链接  │
└─────────────────┘
    │
    ├──────────────┬──────────────┐
    ▼              ▼              ▼
┌────────┐    ┌────────┐    ┌────────┐
│ 确认   │    │ 驳回   │    │ 30分钟 │
│        │    │        │    │ 未响应 │
└───┬────┘    └───┬────┘    └───┬────┘
    │              │              │
    ▼              ▼              ▼
SaaS执行      记录驳回原因   维持旧规则
    │              │              │
    ▼              ▼              ▼
记录日志      等待新指令     等待人类批复

6.2 批量确认机制

{
  "batch_confirmation": {
    "max_items": 50,
    "confirmation_methods": [
      "individual",      // 逐条确认
      "category_batch",  // 按类别批量
      "all_at_once"      // 全部确认
    ],
    "timeout_handling": {
      "non_blocking": "wait_until_approved",
      "blocking": "maintain_old_rules"
    },
    "required_fields": [
      "workflow_instance_id",
      "node_id",
      "suggestion_id",
      "tenant_id"
    ]
  }
}

---

7. 风控约束矩阵 (最终版)

---

8. 记忆块规范

8.1 记忆块格式

{
  "timestamp": "2026-06-14T20:00:00Z",
  "phase": "phase_1_design",
  "version": "1.0",
  "summary": "完成15个Agent节点识别和Schema定义，等待人类审批",
  "decisions": [
    "D1: 采用超兔Xtools 7模块架构作为SaaS基础",
    "D2: 项目商机分析Agent作为核心节点，需人类必须确认",
    "D3: 所有Agent输出使用JSON Schema硬校验",
    "D4: 高风险节点(合同/商机/报价)必须人类确认",
    "D5: 多租户物理隔离，Agent结果按租户隔离",
    "D6: 企业微信作为唯一确认渠道",
    "D7: 本地部署DeepSeek-R1 + Qwen2.5混合模型策略",
    "D8: 评分阈值：高≥70, 中50-69, 低30-49, 放弃<30"
  ],
  "dependencies": [],
  "next_steps": [
    "S1: 用户已回答5个业务问题，设计说明书已完善",
    "S2: 人类审批签字后进入阶段2开发",
    "S3: 阶段2任务：编写Agent节点提示词、构建回归测试、CI流水线"
  ],
  "risks": [
    "R1: 项目商机分析Agent输入字段过多，可能影响响应速度",
    "R2: 需要GPU服务器部署DeepSeek-R1-32B/70B",
    "R3: 15个节点同时开发可能导致资源不足，建议分批实施",
    "R4: 企业微信集成需要CorpID和AgentID配置"
  ]
}

8.2 记忆块管理规则

---

9. 版本历史

---

10. 审批记录

**审批人**: admin@topcentral.cn / 麻一明

**审批方式**: 用户消息回复，回答5个业务问题

**审批时间**: 2026-06-14

**审批结论**:

**阶段1状态**: ✅ 完成，可进入阶段2

---

**文档访问**: http://47.96.253.194/docs/saas-agent-node-design-v1.md

**Git仓库**: https://gitee.com/storm0823/topcentral-mall/

**下次更新**: 阶段2完成后更新为v2.0