2537 lines
32 KiB
Markdown
2537 lines
32 KiB
Markdown
# 企业内部智能知识库平台 PRD
|
||
|
||
**文档版本:** v0.1
|
||
**文档状态:** 需求基线稿 / 待评审
|
||
**产品定位:** API First + Admin Console 的企业级智能知识基础设施
|
||
**目标用户规模:** 企业内部几十人~百人级团队,架构预留更大规模扩展能力
|
||
**主要语言:** 中文优先,兼容英文及中英混合内容
|
||
**部署原则:** 自建部署为主,允许调用云端 AI 服务;支持 NVIDIA GPU
|
||
**核心技术原则:** RAG + Knowledge Compilation、Hybrid Retrieval、Reranker、MCP、底层能力可替换
|
||
|
||
---
|
||
|
||
# 1. 项目背景
|
||
|
||
企业内部存在大量分散于 PDF、Word、PPT、Excel、Markdown、图片、扫描件及其他文件中的知识资产。
|
||
|
||
传统文档管理和 Wiki 系统能够解决知识存放、目录组织和人工检索问题,但面对以下场景能力有限:
|
||
|
||
- 用户不知道准确关键词;
|
||
- 一个问题需要综合多份文档;
|
||
- 文档数量增加后,人工维护成本快速上升;
|
||
- AI Agent 无法以统一协议安全访问企业知识;
|
||
- 文档更新后,知识索引不能自动同步;
|
||
- 搜索结果缺乏来源追踪;
|
||
- 不同部门之间存在严格知识访问权限;
|
||
- 原始资料虽然很多,但没有进一步沉淀成结构化企业知识。
|
||
|
||
本项目计划建设一套企业内部智能知识库平台,将企业原始资料自动转换为:
|
||
|
||
1. **可追溯的结构化知识;**
|
||
2. **可供 RAG 高质量召回的索引;**
|
||
3. **可由 LLM 持续维护的 Compiled Knowledge;**
|
||
4. **可通过 REST API / MCP 被 WorkBuddy 等 Agent 调用的知识能力。**
|
||
|
||
平台不绑定单一 Embedding 模型、Vector Database、LLM 或 Agent 客户端。
|
||
|
||
---
|
||
|
||
# 2. 产品愿景
|
||
|
||
建设企业内部统一的 **Knowledge Infrastructure**:
|
||
|
||
> 任何企业知识进入系统后,都能够自动被解析、清洗、结构化、索引、沉淀和安全调用,并且始终能够追溯到原始事实来源。
|
||
|
||
系统长期形成三层知识:
|
||
|
||
```text
|
||
L0 Source Knowledge
|
||
原始文件 / 原始事实
|
||
不可由 AI 修改
|
||
↓
|
||
L1 Structured Knowledge
|
||
解析文本 / 结构 / Chunk / Index
|
||
用于精准检索与 RAG
|
||
↓
|
||
L2 Compiled Knowledge
|
||
Wiki / FAQ / SOP / Entity / Summary / Comparison
|
||
由 AI 从 L0/L1 派生,可持续维护
|
||
```
|
||
|
||
要求:
|
||
|
||
```text
|
||
L2 → L1 → L0
|
||
```
|
||
|
||
始终可完整追溯。
|
||
|
||
---
|
||
|
||
# 3. 核心产品原则
|
||
|
||
## 3.1 API First
|
||
|
||
所有核心业务能力首先以 Service/API 形式存在。
|
||
|
||
Admin Console、MCP Server 以及未来其他客户端均调用统一业务服务,不允许各自重复实现业务逻辑。
|
||
|
||
```text
|
||
Admin Console
|
||
│
|
||
REST
|
||
│
|
||
WorkBuddy ─ MCP ─ Knowledge Service ─ REST ─ Other Systems
|
||
│
|
||
Core Engine
|
||
```
|
||
|
||
---
|
||
|
||
## 3.2 Admin Console 不是核心业务逻辑
|
||
|
||
Admin Console 是官方管理客户端,主要解决:
|
||
|
||
- 知识库创建;
|
||
- 文件上传;
|
||
- 文件状态管理;
|
||
- 权限管理;
|
||
- 检索调试;
|
||
- Chunk 查看;
|
||
- Knowledge Compilation 管理;
|
||
- 模型配置;
|
||
- Benchmark;
|
||
- MCP 配置;
|
||
- 日志和任务监控。
|
||
|
||
所有能力原则上均能够通过 API 操作。
|
||
|
||
---
|
||
|
||
## 3.3 底层实现不得暴露给 MCP Client
|
||
|
||
WorkBuddy、Cursor、Claude Code 或其他 Agent 不应该知道:
|
||
|
||
- Mochow;
|
||
- Qdrant;
|
||
- Milvus;
|
||
- bge-large-zh;
|
||
- Qwen3;
|
||
- MinerU;
|
||
- 其他 Parser/Model。
|
||
|
||
MCP 只暴露业务语义,例如:
|
||
|
||
```text
|
||
search_knowledge
|
||
get_document
|
||
get_document_section
|
||
get_source
|
||
list_knowledge_bases
|
||
```
|
||
|
||
---
|
||
|
||
## 3.4 Source of Truth 原则
|
||
|
||
原始文件是企业知识的最终事实源。
|
||
|
||
LLM 生成的:
|
||
|
||
- Wiki;
|
||
- FAQ;
|
||
- Summary;
|
||
- Comparison;
|
||
- SOP;
|
||
- Entity Page;
|
||
|
||
均属于 **Derived Knowledge**。
|
||
|
||
Derived Knowledge 不得覆盖、修改或替代 Source Knowledge。
|
||
|
||
---
|
||
|
||
## 3.5 Provider 可替换原则
|
||
|
||
核心系统不得与单一供应商强耦合。
|
||
|
||
至少抽象:
|
||
|
||
```text
|
||
EmbeddingProvider
|
||
RerankerProvider
|
||
VectorStoreProvider
|
||
LLMProvider
|
||
ParserProvider
|
||
OCRProvider
|
||
ObjectStorageProvider
|
||
```
|
||
|
||
V1 可以优先使用现有百度体系,但必须具备替换路径。
|
||
|
||
---
|
||
|
||
# 4. 产品目标
|
||
|
||
## 4.1 V1 核心目标
|
||
|
||
V1 必须实现以下完整闭环:
|
||
|
||
```text
|
||
创建知识库
|
||
↓
|
||
上传文档
|
||
↓
|
||
自动解析
|
||
↓
|
||
自动清洗
|
||
↓
|
||
结构恢复
|
||
↓
|
||
Chunk
|
||
↓
|
||
Embedding
|
||
↓
|
||
Hybrid Index
|
||
↓
|
||
Reranker
|
||
↓
|
||
带权限检索
|
||
↓
|
||
来源追踪
|
||
↓
|
||
REST / MCP
|
||
```
|
||
|
||
同时实现:
|
||
|
||
```text
|
||
Raw Knowledge
|
||
↓
|
||
Knowledge Compilation
|
||
↓
|
||
Compiled Wiki
|
||
↓
|
||
审核 / 发布
|
||
↓
|
||
RAG 可检索
|
||
```
|
||
|
||
---
|
||
|
||
# 5. 非目标
|
||
|
||
以下能力不作为 V1 核心目标:
|
||
|
||
- 自研通用聊天机器人;
|
||
- 完整 Agent 平台;
|
||
- 通用 Workflow 编排平台;
|
||
- GraphRAG;
|
||
- 知识图谱编辑平台;
|
||
- 视频语义理解;
|
||
- 音频会议实时转写;
|
||
- 自动互联网爬虫;
|
||
- 飞书 / Confluence / Notion 全量同步;
|
||
- 多模态图片向量检索;
|
||
- 完整文档在线协同编辑器;
|
||
- 移动端 App。
|
||
|
||
但架构不得阻碍未来扩展。
|
||
|
||
---
|
||
|
||
# 6. 用户角色
|
||
|
||
## 6.1 System Admin
|
||
|
||
系统管理员。
|
||
|
||
权限:
|
||
|
||
- 系统配置;
|
||
- 用户/组织管理;
|
||
- Provider 配置;
|
||
- Embedding / Reranker 管理;
|
||
- MCP 管理;
|
||
- Benchmark;
|
||
- 全局审计;
|
||
- 系统日志;
|
||
- 存储配置。
|
||
|
||
---
|
||
|
||
## 6.2 Knowledge Base Owner
|
||
|
||
知识库负责人。
|
||
|
||
权限:
|
||
|
||
- 创建/配置知识库;
|
||
- 成员管理;
|
||
- 文件上传;
|
||
- 文件删除;
|
||
- 权限设置;
|
||
- Compilation 配置;
|
||
- 发布 Compiled Knowledge;
|
||
- 查看日志;
|
||
- 查看检索效果。
|
||
|
||
---
|
||
|
||
## 6.3 Knowledge Editor
|
||
|
||
知识维护人员。
|
||
|
||
权限:
|
||
|
||
- 上传文档;
|
||
- 编辑 metadata;
|
||
- 查看解析结果;
|
||
- 重新解析;
|
||
- 管理 Compiled Knowledge;
|
||
- 发起发布审核。
|
||
|
||
---
|
||
|
||
## 6.4 Knowledge Consumer
|
||
|
||
普通知识消费者。
|
||
|
||
权限:
|
||
|
||
- 检索其有权限访问的知识;
|
||
- 查看来源;
|
||
- 通过授权 MCP Client 获取知识。
|
||
|
||
不能访问:
|
||
|
||
- 无权限知识库;
|
||
- 无权限文档;
|
||
- 系统管理配置。
|
||
|
||
---
|
||
|
||
## 6.5 AI Agent
|
||
|
||
包括:
|
||
|
||
- WorkBuddy;
|
||
- CodeBuddy;
|
||
- Cursor;
|
||
- Claude Code;
|
||
- 其他 MCP Client;
|
||
- 企业内部 Agent。
|
||
|
||
Agent 的实际权限必须绑定用户身份或 Service Account,不得拥有默认全局权限。
|
||
|
||
---
|
||
|
||
# 7. 核心业务对象
|
||
|
||
系统核心对象:
|
||
|
||
```text
|
||
Organization
|
||
User
|
||
Group
|
||
Role
|
||
|
||
KnowledgeBase
|
||
Folder
|
||
Document
|
||
DocumentVersion
|
||
DocumentPage
|
||
DocumentSection
|
||
Chunk
|
||
|
||
EmbeddingIndex
|
||
SparseIndex
|
||
IndexVersion
|
||
|
||
CompiledKnowledge
|
||
CompiledKnowledgeVersion
|
||
CompilationJob
|
||
|
||
Permission
|
||
ACL
|
||
|
||
IngestionJob
|
||
RetrievalRequest
|
||
RetrievalResult
|
||
AuditLog
|
||
MCPClient
|
||
APIKey
|
||
```
|
||
|
||
---
|
||
|
||
# 8. Knowledge Base 模型
|
||
|
||
知识库是逻辑隔离的最上层业务空间。
|
||
|
||
每个知识库至少包含:
|
||
|
||
- Name;
|
||
- Description;
|
||
- Owner;
|
||
- Visibility;
|
||
- Members;
|
||
- ACL;
|
||
- Folder Tree;
|
||
- Documents;
|
||
- Retrieval Configuration;
|
||
- Embedding Configuration;
|
||
- Reranker Configuration;
|
||
- Compilation Configuration;
|
||
- MCP Availability;
|
||
- Metadata Schema。
|
||
|
||
支持知识库状态:
|
||
|
||
```text
|
||
Active
|
||
Archived
|
||
Disabled
|
||
```
|
||
|
||
---
|
||
|
||
# 9. 文件类型需求
|
||
|
||
## 9.1 P0 文件类型
|
||
|
||
V1 必须支持:
|
||
|
||
- PDF;
|
||
- DOCX;
|
||
- PPTX;
|
||
- XLSX;
|
||
- CSV;
|
||
- Markdown;
|
||
- TXT;
|
||
- HTML;
|
||
- JPG/JPEG;
|
||
- PNG。
|
||
|
||
---
|
||
|
||
## 9.2 PDF 类型
|
||
|
||
必须覆盖:
|
||
|
||
- 原生文本 PDF;
|
||
- 扫描 PDF;
|
||
- 图文混排 PDF;
|
||
- 多栏 PDF;
|
||
- 含表格 PDF;
|
||
- 含页眉页脚 PDF。
|
||
|
||
---
|
||
|
||
# 10. 文件上传
|
||
|
||
**REQ-INGEST-001**
|
||
|
||
支持:
|
||
|
||
- 单文件上传;
|
||
- 多文件批量上传;
|
||
- Drag & Drop。
|
||
|
||
**REQ-INGEST-002**
|
||
|
||
上传完成必须立即返回 Document ID,不阻塞等待解析。
|
||
|
||
**REQ-INGEST-003**
|
||
|
||
Document processing 必须异步执行。
|
||
|
||
**REQ-INGEST-004**
|
||
|
||
展示处理阶段:
|
||
|
||
```text
|
||
Uploading
|
||
Queued
|
||
Parsing
|
||
Cleaning
|
||
Chunking
|
||
Embedding
|
||
Indexing
|
||
Completed
|
||
Failed
|
||
```
|
||
|
||
**REQ-INGEST-005**
|
||
|
||
失败必须展示:
|
||
|
||
- Failed Stage;
|
||
- Error Code;
|
||
- Error Message;
|
||
- Retry。
|
||
|
||
---
|
||
|
||
# 11. 文件版本和去重
|
||
|
||
## REQ-VERSION-001
|
||
|
||
上传文件时计算 Source Hash。
|
||
|
||
如果:
|
||
|
||
```text
|
||
相同 Knowledge Base
|
||
+
|
||
相同文件 Hash
|
||
```
|
||
|
||
系统应提示文件已经存在。
|
||
|
||
---
|
||
|
||
## REQ-VERSION-002
|
||
|
||
同名但 Hash 不同的文件允许:
|
||
|
||
```text
|
||
Create New Version
|
||
```
|
||
|
||
---
|
||
|
||
## REQ-VERSION-003
|
||
|
||
新版本成功建立索引后,旧版本默认退出 Active Retrieval。
|
||
|
||
历史版本仍可追踪。
|
||
|
||
---
|
||
|
||
## REQ-VERSION-004
|
||
|
||
禁止出现:
|
||
|
||
```text
|
||
新版本已经上传
|
||
但检索结果仍同时返回多个旧版本
|
||
```
|
||
|
||
除非管理员明确允许历史搜索。
|
||
|
||
---
|
||
|
||
# 12. Document Parsing
|
||
|
||
解析系统必须使用 Parser Adapter。
|
||
|
||
```text
|
||
ParserRouter
|
||
├─ PDFParser
|
||
├─ OfficeParser
|
||
├─ SpreadsheetParser
|
||
├─ MarkdownParser
|
||
├─ HTMLParser
|
||
└─ ImageParser
|
||
```
|
||
|
||
V1 不规定必须使用单一 Parser。
|
||
|
||
MinerU、Docling 等由下一阶段 Benchmark 决定。
|
||
|
||
---
|
||
|
||
# 13. OCR
|
||
|
||
系统必须能够:
|
||
|
||
- 自动检测扫描 PDF;
|
||
- 自动检测图片中的文字;
|
||
- 启动 OCR Pipeline;
|
||
- 标记 OCR confidence;
|
||
- 保存原始页码。
|
||
|
||
对于低置信度 OCR 页面,应在管理后台显示警告。
|
||
|
||
---
|
||
|
||
# 14. Document Cleaning
|
||
|
||
系统必须支持自动清洗:
|
||
|
||
- 空页;
|
||
- 多余空格;
|
||
- 异常换行;
|
||
- 连字符断词;
|
||
- 重复页眉;
|
||
- 重复页脚;
|
||
- 页码;
|
||
- OCR 噪声;
|
||
- 无意义字符;
|
||
- 重复内容。
|
||
|
||
要求:
|
||
|
||
**不得因为清洗破坏原始 Source。**
|
||
|
||
保存:
|
||
|
||
```text
|
||
Original
|
||
Parsed
|
||
Cleaned
|
||
```
|
||
|
||
三个逻辑层次。
|
||
|
||
---
|
||
|
||
# 15. 文档结构恢复
|
||
|
||
解析结果至少需要识别:
|
||
|
||
```text
|
||
Document
|
||
├─ Title
|
||
├─ Heading
|
||
├─ Paragraph
|
||
├─ List
|
||
├─ Table
|
||
├─ Image Reference
|
||
└─ Page
|
||
```
|
||
|
||
每个结构节点至少保留:
|
||
|
||
```text
|
||
source_document_id
|
||
document_version
|
||
page_number
|
||
section_path
|
||
block_type
|
||
content
|
||
```
|
||
|
||
---
|
||
|
||
# 16. Spreadsheet 处理
|
||
|
||
Excel / CSV 不允许简单拼成连续文本。
|
||
|
||
至少支持:
|
||
|
||
- Sheet 识别;
|
||
- Header 识别;
|
||
- Row/Column 信息;
|
||
- Cell 内容;
|
||
- Sheet metadata。
|
||
|
||
Chunk 应保留:
|
||
|
||
```text
|
||
sheet_name
|
||
row_range
|
||
column_range
|
||
```
|
||
|
||
方便来源追踪。
|
||
|
||
---
|
||
|
||
# 17. Chunking
|
||
|
||
Chunk Strategy 必须可配置和扩展。
|
||
|
||
支持:
|
||
|
||
### Structure-aware Chunk
|
||
|
||
优先按照:
|
||
|
||
```text
|
||
Heading
|
||
Paragraph
|
||
List
|
||
Table
|
||
```
|
||
|
||
进行切分。
|
||
|
||
### Token-aware Chunk
|
||
|
||
确保满足当前 Embedding Model 最大输入限制。
|
||
|
||
---
|
||
|
||
## REQ-CHUNK-001
|
||
|
||
Chunk 至少保存:
|
||
|
||
```text
|
||
chunk_id
|
||
document_id
|
||
document_version
|
||
knowledge_base_id
|
||
content
|
||
title
|
||
section_path
|
||
page_start
|
||
page_end
|
||
chunk_index
|
||
content_type
|
||
source_hash
|
||
metadata
|
||
ACL
|
||
```
|
||
|
||
---
|
||
|
||
# 18. Embedding Provider
|
||
|
||
系统必须支持多 Embedding Provider。
|
||
|
||
```text
|
||
EmbeddingProvider
|
||
├─ Baidu Qianfan
|
||
├─ Local Model
|
||
└─ OpenAI-compatible Provider
|
||
```
|
||
|
||
现阶段默认候选:
|
||
|
||
- bge-large-zh;
|
||
- Qwen3-Embedding-0.6B;
|
||
- Qwen3-Embedding-4B;
|
||
- Qwen3-Embedding-8B。
|
||
|
||
千帆当前公开模型能力中,bge-large-zh 为 1024 维、512 tokens 上下文;Qwen3 Embedding 提供 0.6B/4B/8B,对应 1024/2560/4096 维,并支持 8192 tokens 上下文。
|
||
|
||
---
|
||
|
||
# 19. Embedding Benchmark
|
||
|
||
这是 V1 的正式产品能力,而不仅仅是研发脚本。
|
||
|
||
Admin Console 应提供 Benchmark Workspace。
|
||
|
||
---
|
||
|
||
## 19.1 测试集
|
||
|
||
管理员可以创建:
|
||
|
||
```text
|
||
Evaluation Dataset
|
||
```
|
||
|
||
每条数据包含:
|
||
|
||
```text
|
||
Query
|
||
Expected Document
|
||
Expected Section
|
||
Expected Chunk
|
||
Tags
|
||
Difficulty
|
||
```
|
||
|
||
---
|
||
|
||
## 19.2 Benchmark 模型
|
||
|
||
V1 至少比较:
|
||
|
||
```text
|
||
bge-large-zh
|
||
Qwen3-Embedding-0.6B
|
||
Qwen3-Embedding-4B
|
||
```
|
||
|
||
根据 GPU 和成本情况扩展 8B。
|
||
|
||
---
|
||
|
||
## 19.3 Benchmark 指标
|
||
|
||
至少计算:
|
||
|
||
```text
|
||
Recall@1
|
||
Recall@3
|
||
Recall@5
|
||
Recall@10
|
||
MRR
|
||
NDCG@10
|
||
Latency P50
|
||
Latency P95
|
||
Indexing Throughput
|
||
GPU Memory
|
||
API Cost
|
||
Storage Cost
|
||
```
|
||
|
||
---
|
||
|
||
## 19.4 中文测试分类
|
||
|
||
测试集必须覆盖:
|
||
|
||
- 自然语言问题;
|
||
- 精确术语;
|
||
- 产品型号;
|
||
- 编号;
|
||
- 中文简称;
|
||
- 中英文混写;
|
||
- 长问题;
|
||
- 多条件问题;
|
||
- 同义表达;
|
||
- 跨文档问题。
|
||
|
||
---
|
||
|
||
# 20. Vector Store
|
||
|
||
系统采用:
|
||
|
||
```text
|
||
VectorStoreProvider
|
||
```
|
||
|
||
抽象。
|
||
|
||
默认首选:
|
||
|
||
```text
|
||
Mochow
|
||
```
|
||
|
||
原因:
|
||
|
||
公司已有使用经验,同时 Mochow 当前官方能力已经支持向量检索、BM25、Hybrid Search 和标量过滤。
|
||
|
||
但系统不得使用 Mochow-specific data model 污染业务 Domain Model。
|
||
|
||
未来可以新增:
|
||
|
||
```text
|
||
Qdrant
|
||
Milvus
|
||
pgvector
|
||
Elasticsearch/OpenSearch
|
||
```
|
||
|
||
---
|
||
|
||
# 21. Retrieval Pipeline
|
||
|
||
V1 默认检索 Pipeline:
|
||
|
||
```text
|
||
Query
|
||
↓
|
||
Query Normalize
|
||
↓
|
||
ACL Resolution
|
||
↓
|
||
┌───────────────┐
|
||
│ Dense Search │
|
||
└───────────────┘
|
||
+
|
||
┌───────────────┐
|
||
│ Sparse / BM25 │
|
||
└───────────────┘
|
||
↓
|
||
Fusion
|
||
↓
|
||
Candidate Top N
|
||
↓
|
||
Reranker
|
||
↓
|
||
Top K
|
||
↓
|
||
Source Enrichment
|
||
↓
|
||
Response
|
||
```
|
||
|
||
---
|
||
|
||
# 22. Hybrid Search
|
||
|
||
必须支持:
|
||
|
||
- Semantic Search;
|
||
- Keyword Search;
|
||
- Hybrid Search;
|
||
- Metadata Filter;
|
||
- ACL Filter;
|
||
- Knowledge Base Scope;
|
||
- Folder Scope;
|
||
- Document Scope。
|
||
|
||
管理员可以设置默认:
|
||
|
||
```text
|
||
Dense Weight
|
||
Sparse Weight
|
||
Candidate Size
|
||
Final TopK
|
||
```
|
||
|
||
---
|
||
|
||
# 23. Reranker
|
||
|
||
Reranker 是 V1 P0 能力。
|
||
|
||
不允许上线以下单一模式作为默认正式检索链:
|
||
|
||
```text
|
||
Embedding
|
||
→
|
||
TopK
|
||
→
|
||
Return
|
||
```
|
||
|
||
默认应:
|
||
|
||
```text
|
||
Hybrid Recall
|
||
→
|
||
Candidate Top 20~50
|
||
→
|
||
Reranker
|
||
→
|
||
Top 5~10
|
||
```
|
||
|
||
Provider:
|
||
|
||
```text
|
||
RerankerProvider
|
||
```
|
||
|
||
初始候选:
|
||
|
||
- Qwen3-Reranker-0.6B;
|
||
- Qwen3-Reranker-4B;
|
||
- Qwen3-Reranker-8B;
|
||
- bce-reranker-base;
|
||
- Local Reranker。
|
||
|
||
千帆当前已经提供 Qwen3 Reranker 系列,最大输入支持 32K。
|
||
|
||
---
|
||
|
||
# 24. 权限系统
|
||
|
||
权限必须发生在 **Retrieval 阶段之前/期间**。
|
||
|
||
禁止:
|
||
|
||
```text
|
||
全局 Search
|
||
↓
|
||
TopK
|
||
↓
|
||
应用层删除用户无权内容
|
||
```
|
||
|
||
必须:
|
||
|
||
```text
|
||
User
|
||
↓
|
||
Permission Resolver
|
||
↓
|
||
Allowed Scope
|
||
↓
|
||
Retrieval Filter
|
||
↓
|
||
TopK
|
||
```
|
||
|
||
---
|
||
|
||
# 25. ACL 模型
|
||
|
||
至少支持:
|
||
|
||
```text
|
||
Organization
|
||
Knowledge Base
|
||
Folder
|
||
Document
|
||
```
|
||
|
||
权限主体:
|
||
|
||
```text
|
||
User
|
||
Group
|
||
Role
|
||
```
|
||
|
||
权限:
|
||
|
||
```text
|
||
Read
|
||
Write
|
||
Manage
|
||
```
|
||
|
||
支持:
|
||
|
||
```text
|
||
Inherited Permission
|
||
Explicit Permission
|
||
```
|
||
|
||
---
|
||
|
||
# 26. 权限审计
|
||
|
||
所有以下操作记录 Audit Log:
|
||
|
||
- 用户登录;
|
||
- 查看文档;
|
||
- 搜索;
|
||
- 下载 Source;
|
||
- 上传;
|
||
- 删除;
|
||
- 权限修改;
|
||
- MCP 调用;
|
||
- API 调用;
|
||
- Compilation;
|
||
- 发布 Compiled Knowledge;
|
||
- Provider 配置修改。
|
||
|
||
---
|
||
|
||
# 27. Knowledge Compilation
|
||
|
||
Knowledge Compilation 是区别于普通 RAG 系统的核心能力。
|
||
|
||
其思想参考 Karpathy LLM Wiki:
|
||
|
||
> LLM 不仅在每次查询时临时理解知识,而是将原始资料持续编译为可维护、可积累的结构化知识资产。
|
||
|
||
Karpathy 原始 LLM Wiki 方案明确区分 immutable raw source 与由 LLM 持续维护的 wiki。
|
||
|
||
企业版本必须在这一思想之上增加:
|
||
|
||
- 权限;
|
||
- 来源追踪;
|
||
- 版本;
|
||
- 审核;
|
||
- 冲突检测;
|
||
- 自动过期机制。
|
||
|
||
---
|
||
|
||
# 28. Compiled Knowledge 类型
|
||
|
||
V1 支持:
|
||
|
||
```text
|
||
Summary
|
||
Entity
|
||
Concept
|
||
FAQ
|
||
SOP
|
||
Comparison
|
||
Overview
|
||
Custom Wiki Page
|
||
```
|
||
|
||
例如:
|
||
|
||
```text
|
||
产品/
|
||
├─ 产品A.md
|
||
├─ 产品B.md
|
||
└─ 产品对比.md
|
||
|
||
技术/
|
||
├─ 部署流程.md
|
||
├─ 故障排查.md
|
||
└─ 系统架构.md
|
||
|
||
企业制度/
|
||
├─ 报销制度.md
|
||
└─ 请假制度.md
|
||
```
|
||
|
||
---
|
||
|
||
# 29. Compiled Knowledge Source
|
||
|
||
每条 Compiled Knowledge 必须包含:
|
||
|
||
```text
|
||
compiled_id
|
||
title
|
||
type
|
||
content
|
||
|
||
source_document_ids
|
||
source_chunk_ids
|
||
source_document_versions
|
||
|
||
created_by
|
||
created_at
|
||
updated_at
|
||
|
||
status
|
||
permissions
|
||
|
||
last_verified_at
|
||
```
|
||
|
||
---
|
||
|
||
# 30. Knowledge Compilation Workflow
|
||
|
||
```text
|
||
New / Updated Source
|
||
↓
|
||
Impact Analysis
|
||
↓
|
||
Find Related Wiki Pages
|
||
↓
|
||
Generate Proposed Update
|
||
↓
|
||
Source Verification
|
||
↓
|
||
Conflict Detection
|
||
↓
|
||
Draft
|
||
↓
|
||
Human Review / Auto Publish Policy
|
||
↓
|
||
Published
|
||
```
|
||
|
||
---
|
||
|
||
# 31. Compilation 状态
|
||
|
||
```text
|
||
Draft
|
||
ReviewRequired
|
||
Published
|
||
Rejected
|
||
Stale
|
||
Archived
|
||
```
|
||
|
||
V1 默认:
|
||
|
||
> AI 生成内容先进入 Draft/ReviewRequired,不自动覆盖 Published 内容。
|
||
|
||
管理员未来可对低风险知识库开启 Auto Publish。
|
||
|
||
---
|
||
|
||
# 32. Knowledge Conflict Detection
|
||
|
||
例如 Source A:
|
||
|
||
```text
|
||
退款周期:30 天
|
||
```
|
||
|
||
Source B:
|
||
|
||
```text
|
||
退款周期:14 天
|
||
```
|
||
|
||
Compilation 不允许自动生成:
|
||
|
||
```text
|
||
退款周期约 14~30 天
|
||
```
|
||
|
||
必须标记:
|
||
|
||
```text
|
||
Knowledge Conflict
|
||
```
|
||
|
||
并展示:
|
||
|
||
- conflicting claims;
|
||
- source;
|
||
- version;
|
||
- date。
|
||
|
||
---
|
||
|
||
# 33. Stale Detection
|
||
|
||
如果 Compiled Knowledge 引用的 Source:
|
||
|
||
- 被删除;
|
||
- 发布新版本;
|
||
- 权限发生变化;
|
||
|
||
相关 Compiled Knowledge 自动进入:
|
||
|
||
```text
|
||
Stale
|
||
```
|
||
|
||
并触发重新 Compilation。
|
||
|
||
---
|
||
|
||
# 34. Compiled Knowledge 权限
|
||
|
||
默认权限原则:
|
||
|
||
> Compiled Knowledge 的可见范围不能高于其引用 Source 的安全范围。
|
||
|
||
系统必须防止:
|
||
|
||
```text
|
||
Restricted HR Document
|
||
+
|
||
Public Document
|
||
↓
|
||
AI Summary
|
||
↓
|
||
Public Wiki
|
||
```
|
||
|
||
造成权限泄漏。
|
||
|
||
具体权限求交/继承算法在技术设计阶段确定。
|
||
|
||
---
|
||
|
||
# 35. Compiled Knowledge 也进入 Retrieval
|
||
|
||
检索可以有两类 Source:
|
||
|
||
```text
|
||
Raw Structured Knowledge
|
||
Compiled Knowledge
|
||
```
|
||
|
||
Retrieval Response 必须标记:
|
||
|
||
```text
|
||
source_type = RAW
|
||
```
|
||
|
||
或:
|
||
|
||
```text
|
||
source_type = COMPILED
|
||
```
|
||
|
||
Compiled Result 仍必须返回 underlying source。
|
||
|
||
---
|
||
|
||
# 36. Query Routing
|
||
|
||
V1 可先采用规则 + LLM Classifier 判断查询:
|
||
|
||
```text
|
||
Fact Lookup
|
||
Exact Search
|
||
Conceptual Search
|
||
Cross Document Analysis
|
||
```
|
||
|
||
简单事实问题优先 Raw Retrieval。
|
||
|
||
复杂综合问题可:
|
||
|
||
```text
|
||
Compiled Knowledge
|
||
+
|
||
Raw Evidence
|
||
```
|
||
|
||
联合召回。
|
||
|
||
---
|
||
|
||
# 37. 来源追踪
|
||
|
||
任何 Retrieval Result 必须包含:
|
||
|
||
```text
|
||
document_id
|
||
filename
|
||
document_version
|
||
page
|
||
section_path
|
||
chunk_id
|
||
content
|
||
score
|
||
rerank_score
|
||
source_type
|
||
```
|
||
|
||
对于 Compiled Knowledge:
|
||
|
||
```text
|
||
compiled_knowledge_id
|
||
underlying_sources[]
|
||
```
|
||
|
||
---
|
||
|
||
# 38. Admin Console
|
||
|
||
产品参考 iWiki / 企业 Wiki 的知识组织思想,但不复制其具体实现。
|
||
|
||
重点吸收:
|
||
|
||
- Space/Knowledge Base;
|
||
- Folder/Category;
|
||
- Content 生命周期;
|
||
- 权限;
|
||
- 版本;
|
||
- 最近更新;
|
||
- 审核;
|
||
- 内容维护责任人;
|
||
- 搜索与浏览结合。
|
||
|
||
腾讯当前企业知识管理体系也将 iWiki 作为可以查询、编辑和管理的企业知识源,并通过 MCP/Connector 与 Agent 连接。
|
||
|
||
---
|
||
|
||
# 39. Admin Console 一级导航
|
||
|
||
建议:
|
||
|
||
```text
|
||
Overview
|
||
|
||
Knowledge Bases
|
||
|
||
Documents
|
||
|
||
Compiled Knowledge
|
||
|
||
Search Lab
|
||
|
||
Evaluation
|
||
|
||
Models
|
||
|
||
MCP & API
|
||
|
||
Users & Permissions
|
||
|
||
Jobs
|
||
|
||
Audit Logs
|
||
|
||
System Settings
|
||
```
|
||
|
||
---
|
||
|
||
# 40. Knowledge Base 页面
|
||
|
||
包含:
|
||
|
||
```text
|
||
Overview
|
||
Documents
|
||
Folders
|
||
Compiled Knowledge
|
||
Search Test
|
||
Evaluation
|
||
Permissions
|
||
Settings
|
||
```
|
||
|
||
---
|
||
|
||
# 41. Document 页面
|
||
|
||
展示:
|
||
|
||
- 原始文件;
|
||
- 文件信息;
|
||
- 当前版本;
|
||
- 历史版本;
|
||
- Parser;
|
||
- Processing Pipeline;
|
||
- Parsed Result;
|
||
- Chunk;
|
||
- Metadata;
|
||
- Permissions;
|
||
- Related Compiled Knowledge;
|
||
- Index Status。
|
||
|
||
---
|
||
|
||
# 42. Chunk Inspector
|
||
|
||
管理员可以查看:
|
||
|
||
```text
|
||
Chunk #23
|
||
|
||
Document:
|
||
产品管理制度.pdf
|
||
|
||
Page:
|
||
17
|
||
|
||
Section:
|
||
第三章 / 产品退款
|
||
|
||
Text:
|
||
...
|
||
|
||
Embedding Model:
|
||
...
|
||
|
||
Index Version:
|
||
...
|
||
|
||
ACL:
|
||
...
|
||
|
||
Retrieval History:
|
||
...
|
||
```
|
||
|
||
支持:
|
||
|
||
```text
|
||
Re-chunk
|
||
Re-embed
|
||
Re-index
|
||
```
|
||
|
||
---
|
||
|
||
# 43. Search Lab
|
||
|
||
这是 V1 P0 工具。
|
||
|
||
输入 Query 后展示:
|
||
|
||
### Query Information
|
||
|
||
```text
|
||
Normalized Query
|
||
Search Scope
|
||
ACL
|
||
```
|
||
|
||
### Dense Recall
|
||
|
||
```text
|
||
Rank
|
||
Document
|
||
Chunk
|
||
Score
|
||
```
|
||
|
||
### Sparse Recall
|
||
|
||
```text
|
||
Rank
|
||
Document
|
||
Chunk
|
||
BM25 Score
|
||
```
|
||
|
||
### Fusion
|
||
|
||
### Reranker
|
||
|
||
### Final Result
|
||
|
||
通过此功能定位:
|
||
|
||
- Parsing 问题;
|
||
- Chunk 问题;
|
||
- Embedding 问题;
|
||
- BM25 问题;
|
||
- Reranker 问题。
|
||
|
||
---
|
||
|
||
# 44. MCP Server
|
||
|
||
系统提供统一:
|
||
|
||
```text
|
||
Enterprise Knowledge MCP Server
|
||
```
|
||
|
||
不得直接暴露 Vector DB API。
|
||
|
||
---
|
||
|
||
# 45. MCP P0 Tools
|
||
|
||
建议 V1:
|
||
|
||
### search_knowledge
|
||
|
||
```text
|
||
query
|
||
knowledge_base_ids?
|
||
folder_ids?
|
||
top_k?
|
||
filters?
|
||
```
|
||
|
||
返回:
|
||
|
||
```text
|
||
results[]
|
||
sources[]
|
||
```
|
||
|
||
---
|
||
|
||
### get_document
|
||
|
||
获取指定 Document metadata。
|
||
|
||
---
|
||
|
||
### get_document_section
|
||
|
||
获取指定章节/页/Chunk 的连续上下文。
|
||
|
||
---
|
||
|
||
### get_source
|
||
|
||
获取完整来源信息。
|
||
|
||
---
|
||
|
||
### list_knowledge_bases
|
||
|
||
只返回当前用户有权限访问的知识库。
|
||
|
||
---
|
||
|
||
# 46. MCP P1 Tools
|
||
|
||
后续支持:
|
||
|
||
```text
|
||
search_compiled_knowledge
|
||
get_compiled_knowledge
|
||
create_knowledge_draft
|
||
submit_knowledge
|
||
```
|
||
|
||
写操作必须有更严格授权。
|
||
|
||
---
|
||
|
||
# 47. MCP 身份
|
||
|
||
远程 MCP 不得使用一个共享超级管理员 Token。
|
||
|
||
要求:
|
||
|
||
```text
|
||
Client
|
||
+
|
||
User / Service Identity
|
||
+
|
||
Permission
|
||
```
|
||
|
||
映射到系统 ACL。
|
||
|
||
---
|
||
|
||
# 48. MCP Client Compatibility
|
||
|
||
首要目标:
|
||
|
||
```text
|
||
WorkBuddy Enterprise
|
||
```
|
||
|
||
同时尽量兼容标准 MCP Client。
|
||
|
||
WorkBuddy Enterprise 当前官方 MCP 文档已经支持自定义 MCP Server,并提供远程 API Server、权限、安全及 iWiki 等配置示例。
|
||
|
||
---
|
||
|
||
# 49. REST API
|
||
|
||
REST API 为正式公开能力。
|
||
|
||
建议 Version:
|
||
|
||
```text
|
||
/api/v1/
|
||
```
|
||
|
||
核心 Domain:
|
||
|
||
```text
|
||
/auth
|
||
/users
|
||
/groups
|
||
|
||
/knowledge-bases
|
||
/documents
|
||
/compiled-knowledge
|
||
|
||
/search
|
||
/evaluations
|
||
|
||
/jobs
|
||
/audit-logs
|
||
|
||
/mcp
|
||
/providers
|
||
```
|
||
|
||
---
|
||
|
||
# 50. API Authentication
|
||
|
||
V1 支持:
|
||
|
||
```text
|
||
User Token
|
||
Service API Key
|
||
```
|
||
|
||
未来预留:
|
||
|
||
```text
|
||
OIDC
|
||
OAuth2
|
||
LDAP
|
||
企业微信 / AD / SSO
|
||
```
|
||
|
||
---
|
||
|
||
# 51. Search API
|
||
|
||
建议核心接口:
|
||
|
||
```text
|
||
POST /api/v1/search
|
||
```
|
||
|
||
Request:
|
||
|
||
```json
|
||
{
|
||
"query": "...",
|
||
"knowledge_base_ids": [],
|
||
"top_k": 8,
|
||
"retrieval_mode": "hybrid"
|
||
}
|
||
```
|
||
|
||
Response 逻辑结构:
|
||
|
||
```json
|
||
{
|
||
"request_id": "...",
|
||
"results": [],
|
||
"timing": {},
|
||
"retrieval": {}
|
||
}
|
||
```
|
||
|
||
不得在 API Contract 中暴露:
|
||
|
||
```text
|
||
mochow.xxx
|
||
qdrant.xxx
|
||
```
|
||
|
||
---
|
||
|
||
# 52. Provider Management
|
||
|
||
管理员可以配置:
|
||
|
||
### Embedding Provider
|
||
|
||
### Reranker Provider
|
||
|
||
### LLM Provider
|
||
|
||
### Vector Store Provider
|
||
|
||
### Parser Provider
|
||
|
||
每个 Provider 具有:
|
||
|
||
```text
|
||
Name
|
||
Type
|
||
Endpoint
|
||
Credential
|
||
Status
|
||
Capabilities
|
||
Health
|
||
```
|
||
|
||
Secret 不允许明文回显。
|
||
|
||
---
|
||
|
||
# 53. LLM Provider
|
||
|
||
Knowledge Compilation 可调用:
|
||
|
||
- 百度千帆;
|
||
- OpenAI-compatible API;
|
||
- 本地模型;
|
||
- 其他云端模型。
|
||
|
||
系统必须支持:
|
||
|
||
```text
|
||
LLMProvider
|
||
```
|
||
|
||
抽象。
|
||
|
||
---
|
||
|
||
# 54. GPU 使用原则
|
||
|
||
本地 NVIDIA GPU 优先服务于:
|
||
|
||
```text
|
||
OCR / Document Parsing
|
||
Embedding
|
||
Reranker
|
||
```
|
||
|
||
LLM 可以采用:
|
||
|
||
```text
|
||
Cloud First
|
||
Local Optional
|
||
```
|
||
|
||
不强制 V1 本地部署大语言模型。
|
||
|
||
---
|
||
|
||
# 55. 可观测性
|
||
|
||
系统必须记录:
|
||
|
||
## Ingestion Metrics
|
||
|
||
```text
|
||
documents/day
|
||
pages/min
|
||
parse latency
|
||
OCR latency
|
||
failed rate
|
||
```
|
||
|
||
## Retrieval Metrics
|
||
|
||
```text
|
||
QPS
|
||
P50
|
||
P95
|
||
P99
|
||
Dense Latency
|
||
Sparse Latency
|
||
Rerank Latency
|
||
```
|
||
|
||
## Quality Metrics
|
||
|
||
```text
|
||
Recall@K
|
||
MRR
|
||
NDCG
|
||
No Result Rate
|
||
```
|
||
|
||
## Compilation Metrics
|
||
|
||
```text
|
||
Jobs
|
||
Success Rate
|
||
Review Rate
|
||
Conflict Rate
|
||
Stale Count
|
||
```
|
||
|
||
---
|
||
|
||
# 56. Job System
|
||
|
||
异步任务至少包括:
|
||
|
||
```text
|
||
ParseJob
|
||
OCRJob
|
||
ChunkJob
|
||
EmbeddingJob
|
||
IndexJob
|
||
CompilationJob
|
||
ReindexJob
|
||
EvaluationJob
|
||
```
|
||
|
||
每个 Job:
|
||
|
||
```text
|
||
Queued
|
||
Running
|
||
Succeeded
|
||
Failed
|
||
Cancelled
|
||
```
|
||
|
||
必须支持:
|
||
|
||
```text
|
||
Retry
|
||
```
|
||
|
||
---
|
||
|
||
# 57. 删除语义
|
||
|
||
删除 Document 时必须处理:
|
||
|
||
```text
|
||
Source
|
||
Parsed Data
|
||
Chunks
|
||
Vector Index
|
||
Sparse Index
|
||
Compiled Knowledge Dependency
|
||
```
|
||
|
||
默认采用:
|
||
|
||
```text
|
||
Soft Delete
|
||
```
|
||
|
||
管理员可执行永久删除。
|
||
|
||
---
|
||
|
||
# 58. 搜索质量 Benchmark
|
||
|
||
正式上线前必须建立企业自己的 Golden Dataset。
|
||
|
||
建议首批不少于:
|
||
|
||
```text
|
||
200~500 Queries
|
||
```
|
||
|
||
随着使用积累扩展。
|
||
|
||
样本来源:
|
||
|
||
- 用户真实问题;
|
||
- 产品团队;
|
||
- HR;
|
||
- 技术团队;
|
||
- 销售/客服。
|
||
|
||
---
|
||
|
||
# 59. Retrieval Benchmark Matrix
|
||
|
||
至少比较:
|
||
|
||
```text
|
||
Embedding × Retrieval × Reranker
|
||
```
|
||
|
||
例如:
|
||
|
||
```text
|
||
bge-large-zh
|
||
+ Mochow Dense
|
||
+ No Reranker
|
||
|
||
bge-large-zh
|
||
+ Mochow Hybrid
|
||
+ Qwen3 Reranker
|
||
|
||
Qwen3 Embedding
|
||
+ Mochow Hybrid
|
||
+ Qwen3 Reranker
|
||
```
|
||
|
||
不只 Benchmark Embedding 单模型。
|
||
|
||
---
|
||
|
||
# 60. Benchmark 决策原则
|
||
|
||
最终生产方案根据以下综合指标选择:
|
||
|
||
```text
|
||
Quality
|
||
Latency
|
||
Cost
|
||
Operational Complexity
|
||
Vendor Dependency
|
||
GPU Requirement
|
||
```
|
||
|
||
而不是只按模型排行榜。
|
||
|
||
---
|
||
|
||
# 61. 安全要求
|
||
|
||
P0:
|
||
|
||
- 数据逻辑隔离;
|
||
- ACL;
|
||
- Secret 加密;
|
||
- API Authentication;
|
||
- Audit Log;
|
||
- MCP Authorization;
|
||
- HTTPS;
|
||
- 禁止未经授权下载原文件。
|
||
|
||
---
|
||
|
||
# 62. 云端 AI 数据安全
|
||
|
||
因为允许调用云端服务,每个 Provider 必须定义:
|
||
|
||
```text
|
||
Data Processing Policy
|
||
```
|
||
|
||
管理员可以设置:
|
||
|
||
```text
|
||
Cloud Allowed
|
||
Local Only
|
||
```
|
||
|
||
Knowledge Base 可覆盖系统默认策略。
|
||
|
||
例如 HR Knowledge Base 可以配置:
|
||
|
||
```text
|
||
Local Only
|
||
```
|
||
|
||
即不得将文档内容发送给任何外部 LLM/Embedding API。
|
||
|
||
---
|
||
|
||
# 63. 性能目标
|
||
|
||
按当前几十~百人团队规模,V1 目标:
|
||
|
||
### Search
|
||
|
||
正常场景:
|
||
|
||
```text
|
||
P95 < 2s
|
||
```
|
||
|
||
包含 Reranker。
|
||
|
||
不包含最终 LLM Answer Generation。
|
||
|
||
### API
|
||
|
||
Metadata API:
|
||
|
||
```text
|
||
P95 < 500ms
|
||
```
|
||
|
||
### Admin Console
|
||
|
||
主要页面交互:
|
||
|
||
```text
|
||
P95 < 2s
|
||
```
|
||
|
||
---
|
||
|
||
# 64. 可用性
|
||
|
||
V1 内部生产目标:
|
||
|
||
```text
|
||
99.5%
|
||
```
|
||
|
||
后续视业务重要性提升。
|
||
|
||
---
|
||
|
||
# 65. 数据规模
|
||
|
||
当前尚未明确文档总量,因此 V1 架构最低要求:
|
||
|
||
```text
|
||
10^6 Chunk 级别
|
||
```
|
||
|
||
不需要大规模重构即可运行。
|
||
|
||
更高规模在架构设计阶段进行 Capacity Planning。
|
||
|
||
---
|
||
|
||
# 66. iWiki 产品设计参考原则
|
||
|
||
不是复制 iWiki,而是吸收传统企业知识系统已经验证的产品理念:
|
||
|
||
### 知识是长期资产
|
||
|
||
不能只有“上传文件”。
|
||
|
||
需要:
|
||
|
||
```text
|
||
Owner
|
||
Version
|
||
History
|
||
Category
|
||
Permission
|
||
Status
|
||
Last Updated
|
||
```
|
||
|
||
### Search + Browse
|
||
|
||
用户不仅通过搜索进入知识,也需要:
|
||
|
||
```text
|
||
Knowledge Base
|
||
→ Folder
|
||
→ Document
|
||
→ Knowledge
|
||
```
|
||
|
||
浏览。
|
||
|
||
### Knowledge Lifecycle
|
||
|
||
```text
|
||
Create
|
||
Maintain
|
||
Review
|
||
Publish
|
||
Update
|
||
Archive
|
||
```
|
||
|
||
### AI 不替代治理
|
||
|
||
AI 可以:
|
||
|
||
```text
|
||
Generate
|
||
Summarize
|
||
Compile
|
||
Detect
|
||
Suggest
|
||
```
|
||
|
||
但企业重要知识需要:
|
||
|
||
```text
|
||
Review
|
||
Ownership
|
||
Audit
|
||
```
|
||
|
||
---
|
||
|
||
# 67. RAG + Knowledge Compilation 的产品定位
|
||
|
||
RAG 负责:
|
||
|
||
> 找到正确事实。
|
||
|
||
Knowledge Compilation 负责:
|
||
|
||
> 让企业知识持续沉淀并减少重复理解成本。
|
||
|
||
二者不存在替代关系。
|
||
|
||
最终查询:
|
||
|
||
```text
|
||
User Query
|
||
│
|
||
Query Router
|
||
/ \
|
||
↓ ↓
|
||
Raw Retrieval Compiled Retrieval
|
||
↓ ↓
|
||
\ /
|
||
Reranker
|
||
↓
|
||
Evidence
|
||
↓
|
||
MCP / REST / Agent
|
||
```
|
||
|
||
---
|
||
|
||
# 68. 核心成功指标
|
||
|
||
V1 上线后重点观察:
|
||
|
||
### Adoption
|
||
|
||
```text
|
||
WAU
|
||
Active Knowledge Bases
|
||
Search Users
|
||
MCP Users
|
||
```
|
||
|
||
### Knowledge Coverage
|
||
|
||
```text
|
||
Indexed Documents %
|
||
Successful Parse %
|
||
```
|
||
|
||
### Search Quality
|
||
|
||
```text
|
||
Recall@5
|
||
Recall@10
|
||
No Result Rate
|
||
User Positive Feedback
|
||
```
|
||
|
||
### Knowledge Maintenance
|
||
|
||
```text
|
||
Compiled Knowledge Count
|
||
Stale Rate
|
||
Conflict Rate
|
||
Review Completion Rate
|
||
```
|
||
|
||
### Reliability
|
||
|
||
```text
|
||
Index Failure Rate
|
||
API Error Rate
|
||
MCP Error Rate
|
||
```
|
||
|
||
---
|
||
|
||
# 69. V1 P0 Scope
|
||
|
||
必须上线:
|
||
|
||
1. 用户与权限;
|
||
2. Knowledge Base;
|
||
3. Folder;
|
||
4. 文件上传;
|
||
5. PDF/DOCX/PPTX/XLSX/CSV/MD/TXT/HTML/Image;
|
||
6. OCR;
|
||
7. Parsing;
|
||
8. Cleaning;
|
||
9. Structure-aware Chunk;
|
||
10. Document Version;
|
||
11. Hash 去重;
|
||
12. Embedding Provider;
|
||
13. Mochow Provider;
|
||
14. Dense Search;
|
||
15. BM25;
|
||
16. Hybrid Retrieval;
|
||
17. ACL Filter;
|
||
18. Reranker;
|
||
19. Source Tracking;
|
||
20. Search API;
|
||
21. MCP Server;
|
||
22. WorkBuddy 接入;
|
||
23. Admin Console;
|
||
24. Search Lab;
|
||
25. Job 管理;
|
||
26. Audit Log;
|
||
27. Knowledge Compilation;
|
||
28. Compiled Knowledge 审核;
|
||
29. Benchmark;
|
||
30. Provider 配置。
|
||
|
||
---
|
||
|
||
# 70. V1.1 / P1
|
||
|
||
后续优先:
|
||
|
||
- LDAP/OIDC;
|
||
- AD/企业 SSO;
|
||
- URL Import;
|
||
- Confluence;
|
||
- iWiki Import;
|
||
- 飞书;
|
||
- 企业微信;
|
||
- Scheduled Sync;
|
||
- MCP 写入知识;
|
||
- Auto Compilation Policy;
|
||
- Feedback Learning;
|
||
- Query Analytics。
|
||
|
||
---
|
||
|
||
# 71. P2
|
||
|
||
未来:
|
||
|
||
- GraphRAG;
|
||
- Knowledge Graph;
|
||
- 多模态 Retrieval;
|
||
- 图片 Embedding;
|
||
- 视频知识;
|
||
- Audio;
|
||
- Agent Workflow;
|
||
- Enterprise Search;
|
||
- Personalized Retrieval;
|
||
- Automatic Knowledge Gap Discovery。
|
||
|
||
---
|
||
|
||
# 72. V1 验收标准
|
||
|
||
产品只有同时满足以下条件才认为 V1 完成。
|
||
|
||
## Ingestion
|
||
|
||
上传 P0 文件类型能够:
|
||
|
||
```text
|
||
Upload → Parse → Clean → Chunk → Embed → Index
|
||
```
|
||
|
||
全链路完成。
|
||
|
||
---
|
||
|
||
## Retrieval
|
||
|
||
支持:
|
||
|
||
```text
|
||
Dense
|
||
BM25
|
||
Hybrid
|
||
Reranker
|
||
```
|
||
|
||
并通过 Admin Search Lab 查看每个阶段。
|
||
|
||
---
|
||
|
||
## Permission
|
||
|
||
用户不能:
|
||
|
||
- 搜索;
|
||
- 浏览;
|
||
- API;
|
||
- MCP;
|
||
|
||
访问无权限 Document。
|
||
|
||
---
|
||
|
||
## Traceability
|
||
|
||
每条 Search Result 可以定位:
|
||
|
||
```text
|
||
Source File
|
||
Version
|
||
Page
|
||
Section
|
||
Chunk
|
||
```
|
||
|
||
---
|
||
|
||
## Model
|
||
|
||
Embedding 和 Reranker 可以通过 Provider 配置替换。
|
||
|
||
---
|
||
|
||
## Vector DB
|
||
|
||
Mochow 可以作为默认生产 Vector Store。
|
||
|
||
业务代码不存在必须依赖 Mochow SDK 数据结构的 API Contract。
|
||
|
||
---
|
||
|
||
## Benchmark
|
||
|
||
能够用同一个 Evaluation Dataset 对多个 Embedding Pipeline 进行比较。
|
||
|
||
---
|
||
|
||
## Knowledge Compilation
|
||
|
||
能够:
|
||
|
||
```text
|
||
Source
|
||
→
|
||
Generate Draft Knowledge
|
||
→
|
||
Source Linking
|
||
→
|
||
Review
|
||
→
|
||
Publish
|
||
```
|
||
|
||
并能够在 Source 更新时把关联知识标记为 Stale。
|
||
|
||
---
|
||
|
||
## MCP
|
||
|
||
WorkBuddy 能通过 MCP:
|
||
|
||
```text
|
||
search_knowledge
|
||
get_source
|
||
```
|
||
|
||
取得与 REST API 一致、且经过 ACL 的结果。
|
||
|
||
---
|
||
|
||
# 73. 当前技术基线
|
||
|
||
PRD 阶段暂定但不最终锁定:
|
||
|
||
```text
|
||
Primary Vector Store:
|
||
Mochow
|
||
|
||
Embedding Baseline:
|
||
bge-large-zh
|
||
|
||
Embedding Benchmark:
|
||
Qwen3 Embedding
|
||
|
||
Reranker:
|
||
Qwen3 Reranker / equivalent
|
||
|
||
Document Parsing:
|
||
Benchmark MinerU / Docling
|
||
|
||
LLM:
|
||
Provider abstraction
|
||
|
||
MCP:
|
||
Standard MCP Server
|
||
|
||
Frontend:
|
||
Admin Console
|
||
|
||
Backend:
|
||
API First
|
||
```
|
||
|
||
---
|
||
|
||
# 74. 已确认产品决策
|
||
|
||
当前已经确认:
|
||
|
||
**DEC-001**
|
||
|
||
采用:
|
||
|
||
```text
|
||
API First + Admin Console
|
||
```
|
||
|
||
---
|
||
|
||
**DEC-002**
|
||
|
||
Mochow 继续作为生产首选候选,同时必须保留 Vector Store 替换能力。
|
||
|
||
---
|
||
|
||
**DEC-003**
|
||
|
||
不直接确定单一 Embedding 模型。
|
||
|
||
必须进行多模型企业中文 Benchmark。
|
||
|
||
---
|
||
|
||
**DEC-004**
|
||
|
||
Reranker 是正式检索链 P0 能力。
|
||
|
||
---
|
||
|
||
**DEC-005**
|
||
|
||
产品设计参考 iWiki / 企业 Wiki 的知识组织、权限和知识生命周期思路。
|
||
|
||
---
|
||
|
||
**DEC-006**
|
||
|
||
采用:
|
||
|
||
```text
|
||
RAG + Knowledge Compilation
|
||
```
|
||
|
||
而不是二选一。
|
||
|
||
---
|
||
|
||
**DEC-007**
|
||
|
||
MCP 只能暴露业务能力,禁止泄露 Mochow、Embedding Model 等底层实现。
|
||
|
||
---
|
||
|
||
**DEC-008**
|
||
|
||
企业知识必须始终能够追溯至 immutable Source。
|
||
|
||
---
|
||
|
||
# 75. 下一阶段需要解决的架构问题
|
||
|
||
PRD 评审通过后进入《总体技术架构设计》。
|
||
|
||
重点决策:
|
||
|
||
1. Backend 技术栈;
|
||
2. Admin Console 技术栈;
|
||
3. PostgreSQL 数据模型;
|
||
4. Object Storage;
|
||
5. Task Queue;
|
||
6. MinerU vs Docling;
|
||
7. OCR Pipeline;
|
||
8. Chunk Strategy;
|
||
9. Mochow Schema;
|
||
10. VectorStore Adapter;
|
||
11. Embedding Provider;
|
||
12. Reranker Serving;
|
||
13. GPU Resource Scheduling;
|
||
14. Knowledge Compilation Engine;
|
||
15. Permission Resolver;
|
||
16. MCP Server;
|
||
17. Authentication;
|
||
18. Deployment Topology;
|
||
19. Observability;
|
||
20. Backup / Disaster Recovery。
|
||
|
||
---
|
||
|
||
# 76. 下一阶段应输出的技术架构方案
|
||
|
||
技术方案至少给出三套:
|
||
|
||
## 方案 A:百度云能力优先
|
||
|
||
```text
|
||
Qianfan
|
||
+
|
||
Mochow
|
||
+
|
||
Local Parsing
|
||
```
|
||
|
||
优点:
|
||
|
||
部署简单、延续现有经验。
|
||
|
||
---
|
||
|
||
## 方案 B:高度自建
|
||
|
||
```text
|
||
Local Embedding
|
||
+
|
||
Local Reranker
|
||
+
|
||
Open-source Vector DB
|
||
+
|
||
Local Parsing
|
||
```
|
||
|
||
优点:
|
||
|
||
数据可控、供应商依赖最低。
|
||
|
||
---
|
||
|
||
## 方案 C:Hybrid 推荐方案
|
||
|
||
```text
|
||
Local Parsing / GPU
|
||
+
|
||
Replaceable Embedding
|
||
+
|
||
Replaceable Reranker
|
||
+
|
||
Mochow Primary
|
||
+
|
||
Cloud LLM
|
||
```
|
||
|
||
目标:
|
||
|
||
在开发成本、搜索质量、隐私、运维成本和供应商锁定之间取得平衡。
|
||
|
||
---
|
||
|
||
# 77. 调研依据备注
|
||
|
||
当前百度千帆模型列表表明 bge-large-zh 与 Qwen3 Embedding、Qwen3 Reranker 可以作为同平台 Benchmark 候选,因此早期无需为模型实验同时引入多个云厂商。
|
||
|
||
Mochow 当前官方 API 已支持向量检索与 BM25 全文检索组合的 Hybrid Search,同时可施加标量 Filter,与企业知识库的 Dense + Sparse + ACL Retrieval 需求相匹配。
|
||
|
||
WorkBuddy Enterprise 当前公开 MCP 文档支持自定义 MCP Server,并提供传输、权限、安全、远程 API Server 以及 iWiki 配置示例,因此“企业知识服务 → MCP → WorkBuddy”的产品路线具备正式产品支持基础。
|
||
|
||
Karpathy 的 LLM Wiki 原始思路强调 immutable raw sources 与由 LLM 持续维护的 compiled wiki 分层,这与本产品提出的 L0 Source / L1 Structured / L2 Compiled Knowledge 模型一致;企业版本在此基础上增加 ACL、版本、审计、冲突和审核机制。
|
||
|
||
---
|
||
|
||
**PRD v0.1 END** |