yawan
f7fc992f57
Anydoc / anydoc parser engine (push) Canceled after 0s
App / Format, vet, test, and build (push) Canceled after 0s
CLI / test (macos-latest / go 1.26) (push) Canceled after 0s
CLI / test (ubuntu-latest / go 1.26) (push) Canceled after 0s
CLI / test (windows-latest / go 1.26) (push) Canceled after 0s
Build and Push Docker Image / build-ui (push) Canceled after 0s
Build and Push Docker Image / build-docreader (push) Canceled after 0s
Build and Push Docker Image / build-sandbox (push) Canceled after 0s
Build and Push Docker Image / build-app (amd64, linux/amd64, ubuntu-latest) (push) Canceled after 0s
Build and Push Docker Image / build-app (arm64, linux/arm64, ubuntu-24.04-arm) (push) Canceled after 0s
DocReader / Python tests and Go client tests (push) Canceled after 0s
Build & Publish (dsh-weknora) / Typecheck, test, pack (push) Canceled after 0s
Build & Publish (dsh-weknora) / End-to-end inside dsh (pinned) (push) Canceled after 0s
Build & Publish (dsh-weknora) / End-to-end inside dsh (latest, advisory) (push) Canceled after 0s
Frontend / Test, type-check, and build (push) Canceled after 0s
Go Lint Cache / warm lint cache (.) (push) Canceled after 0s
Go Lint Cache / warm lint cache (cli) (push) Canceled after 0s
Go Lint Cache / warm lint cache (client) (push) Canceled after 0s
Build & Publish (mcp-server) / Test (Python 3.10) (push) Canceled after 0s
Build & Publish (mcp-server) / Test (Python 3.11) (push) Canceled after 0s
Build & Publish (mcp-server) / Test (Python 3.12) (push) Canceled after 0s
Build & Publish (mcp-server) / Test (Python 3.13) (push) Canceled after 0s
Build and Push Docker Image / merge (push) Canceled after 0s
Build & Publish (dsh-weknora) / Publish to npm (push) Canceled after 0s
Build & Publish (mcp-server) / Build sdist + wheel (push) Canceled after 0s
Build & Publish (mcp-server) / Publish to PyPI (push) Canceled after 0s
96 lines
3.6 KiB
Markdown
96 lines
3.6 KiB
Markdown
# 技术方案 · 售后知识库 POC
|
||
|
||
**项目代号**：Nova-KB-POC
|
||
**版本**：0.3 草案
|
||
**作者**：张明（研发部）
|
||
**评审人**：陈浩（产品部）、赵磊（测试部）
|
||
**最后更新**：2024-01-18
|
||
|
||
## 1. 背景与目标
|
||
|
||
星云科技售后团队每日处理约 400 条安装与排障工单。现有方案依赖 Confluence 全文搜索，一线工程师反馈「搜得到段落但不知道哪句能直接回复客户」。
|
||
|
||
本 POC 目标：
|
||
|
||
1. 导入产品手册、会议纪要、FAQ 等内部文档；
|
||
2. 支持自然语言提问，返回答案 + **可追溯出处**；
|
||
3. 4 周内完成内网演示，供 Q1 规划会中期评审。
|
||
|
||
## 2. 系统架构
|
||
|
||
```text
|
||
[语料来源] Confluence 导出 / Markdown / PDF 手册
|
||
↓
|
||
[WeKnora] docreader 解析 → 分块 → 向量 + BM25 混合检索
|
||
↓
|
||
[模型层] Embedding：本地 bge-m3
|
||
LLM：公司 API 网关 → DeepSeek-V3
|
||
↓
|
||
[接入] 售后工单系统 iframe 嵌入（二期）
|
||
```
|
||
|
||
### 2.1 关键角色
|
||
|
||
| 角色 | 姓名 | 职责 |
|
||
| --- | --- | --- |
|
||
| 技术负责人 | 张明 | 部署、语料管道、检索调优 |
|
||
| 产品对接 | 李薇 | 需求优先级、验收题库 |
|
||
| 测试负责人 | 赵磊 | 标注 100 道验收题、准确率统计 |
|
||
| 审批人 | 陈浩 | 资源申请、范围裁剪 |
|
||
|
||
### 2.2 与中控产品的关系
|
||
|
||
POC 首期语料以 **智能家居中控 Pro** 为主（参见《产品手册》），辅以《Q1 产品会议纪要》中的里程碑与责任人信息，便于回答「谁负责」「什么时候交付」类问题。
|
||
|
||
## 3. 语料清单（首批）
|
||
|
||
| 序号 | 文档 | 格式 | 预估分块 |
|
||
| --- | --- | --- | --- |
|
||
| 1 | 智能家居中控 Pro 产品手册 | Markdown / PDF | 40 |
|
||
| 2 | Q1 产品会议纪要 | Markdown | 25 |
|
||
| 3 | 员工手册 · 报销与休假（售后出差场景） | Markdown | 30 |
|
||
| 4 | Top 50 工单 FAQ | Excel 导入 | 50 |
|
||
|
||
**合计约 145 个知识单元**，满足 3 月 1 日内网演示最低要求。
|
||
|
||
## 4. 检索策略
|
||
|
||
- **默认**：向量 + BM25 混合检索，RRF 融合，Top-5 送入 LLM。
|
||
- **可选**：开启 Rerank（`bge-reranker-v2-m3`）提升多义词场景准确率。
|
||
- **暂不启用**：知识图谱（Neo4j 尚未纳入 POC 环境）、Wiki 自动生成。
|
||
|
||
### 4.1 分块参数（建议）
|
||
|
||
| 参数 | 值 | 说明 |
|
||
| --- | --- | --- |
|
||
| chunk_size | 512 | 与手册段落长度匹配 |
|
||
| chunk_overlap | 64 | 保留表格上下文 |
|
||
| 父子分块 | 开启 | 检索命中子块、生成用父块 |
|
||
|
||
## 5. 验收用例（示例）
|
||
|
||
| 问题 | 期望答案要点 | 期望出处 |
|
||
| --- | --- | --- |
|
||
| 中控 Pro 最多能接多少设备？ | 256，推荐 ≤ 120 | 产品手册 · 技术规格 |
|
||
| Q1 知识库 POC 什么时候验收？ | 2024-03-01 内网上线 | Q1 会议纪要 |
|
||
| 一线城市住宿报销上限？ | 600 元 / 晚 | 员工手册 · 差旅报销 |
|
||
| Matter 认证目标版本？ | 固件 3.5，3 月底灰度 | Q1 会议纪要 |
|
||
| 谁负责售后知识库 POC？ | 张明 | 会议纪要 / 本方案 |
|
||
|
||
赵磊将据此扩展为 100 道正式验收题。
|
||
|
||
## 6. 风险
|
||
|
||
1. **语料过期**：产品手册 v2.1 与即将发布的 v2.2 冲突 → 建立「语料 Owner」机制，李薇每月核对。
|
||
2. **幻觉**：强制 Prompt 要求「仅根据引用回答；无依据则回复不知道」。
|
||
3. **权限**：POC 阶段单租户；生产需按区域售后组拆分知识库 ACL。
|
||
|
||
## 7. 里程碑
|
||
|
||
| 日期 | 交付物 |
|
||
| --- | --- |
|
||
| 2024-01-20 | WeKnora 测试环境可用，导入首批 3 份 Markdown |
|
||
| 2024-02-01 | FAQ Excel 导入完成，混合检索 + Rerank 调参 |
|
||
| 2024-02-05 | 中期评审演示 |
|
||
| 2024-03-01 | 100 题验收报告 |
|