Context for AI

Appearance

The enterprise knowledge workflow for engineering

Turn content from every source into an index and evidence chain for agents, bots and people

Install the Context skills, and your agent knows how to keep knowledge

Production, updates and retrieval are each carried by open skills

Production

Sources become traceable knowledge

Plancontext-planResearch the sources and split the writing workCapturecontextBring code, documents, notes and sessions into the workspaceWriteIndexer skillsWrite articles with citations, per source typeReviewcontextPrepare a diff report for people to review
BuildcontextBuild the knowledge package and websiteArchivecontextRecord the version and changelog

Updates

Only what a change touches is revised

Assesscontext-planCompare source versions and find the affected articlesReviseIndexer skillsRewrite only the affected sectionsReviewcontextPrepare a diff report for people to review
PublishcontextRebuild and publish

Retrieval

Search the knowledge, then reason from the evidence

Askcontext-inspect-searchTake the questionUnderstandcontext-inspect-searchWork out what knowledge the question needsSearchcontext-inspect-searchSearch the knowledge for the relevant articles
Reasoncontext-inspect-searchFollow the evidence chain to the original at a fixed version and reasonAnswercontext-inspect-searchAnswer with sources attached

Build a knowledge service on any agent

Bring the Context workflow into the agents you already run and give them knowledge your team can rely on

Your service agent

Claude CodeCodexCursorManaged agent platform

Bot prompt
Context skill plugin

ProductionUpdatesRetrieval

Remote searchPublishing
  • Knowledge package
  • Knowledge website
  • Git workspace

On the same real questions: more accurate, faster, cheaper

Real developer questions from a large team and production monorepo, against a plain LLM Wiki setup with the code repositoriesHow it was measured →

Fully resolved

LLM Wiki19Context23

+16.7 points

Mean time

LLM Wiki223.2sContext93.2s

58.2% faster

Tool calls

LLM Wiki77.5Context7.8

89.9% fewer

Cost per question

LLM Wiki$2.45Context$0.27

88.9% lower

01

More accurate: knowledge as an index

Indexer skills extract content from every source and stack into a compact knowledge index, and record its evidence chainKnowledge for agents →

Indexers extract what matters

orders/ · order service

  • api/routes.tsAPIs and routes
  • service/refund.tsRules and call paths
  • auth/policy.tsPermission checks
  • config/flags.tsFlags and defaults
  • gen/*.pb.tsGenerated code
  • styles/Styles

IndexerKnowledge: refund flow

Compression

1/115

Sources≈ 202 MB
5.84M lines · 1,129 documents
Knowledge≈ 1.75 MB
220 reviewed articles

Evidence chain →

knowledge / refund flow

  1. Refunds need a paid order within 30 days.1

    service/refund.ts:42-58@8f3c2a1

  2. Refunds over $500 need a manager's approval.2

    auth/policy.ts:17-24@8f3c2a1

  3. Merchants are notified of the result by webhook.3

    Runbook §3v2.4

Every claim records its source: file, lines and version

02

Faster and cheaper: knowledge first, then the original remotely

Knowledge gives the big picture fast; the evidence chain leads remotely to the original for precise reasoningPinned search and reads →

Two routes for one question

Agent sandbox · nothing clonedKnowledge is enough · answer directlySource Graph remote searchEvidence chainKnowledge repoKnowledge repoCode repo @commitCode repo @commitAskAskRead knowledgeRead knowledgeRead originalRead originalAnswerAnswer

Time per query

Local rg10–740 ms
Local git grep9–312 ms
Remote search0.9–2.1 ms

Server side; under 100 ms including the round trip within the same data center

Versions
Several versions of a repository side by side, queried by commit
Same hits
Results match a local search

企业级研发知识工作流

把不同来源的内容整理为索引与证据链，交给 Agent、Bot 和团队使用

装上 Context 技能，Agent 就会维护知识

生产、更新、检索三条流程都由开放的技能完成

生产

把来源写成可追溯的知识

规划context-plan调研来源，拆分写作任务采集context把代码、文档、笔记和会话收进工作区写作Indexer 技能按来源类型写出带引用的文章审核context生成 Diff 报告，交给人审核
构建context构建知识包和知识网站归档context记录版本与变更日志

更新

只修订变化涉及的部分

评估影响context-plan比对来源版本，找出受影响的文章修订Indexer 技能只重写受影响的段落审核context生成 Diff 报告，交给人审核
发布context重新构建并发布

检索

检索知识，再沿证据推理作答

提问context-inspect-search接收问题理解context-inspect-search理解问题，判断需要哪些知识检索context-inspect-search在知识中检索相关文章
推理context-inspect-search沿证据链读取固定版本的原文并推理作答context-inspect-search附上出处作答

在任意 Agent 上搭建知识服务

引入 Context 工作流，增强已有 Agent 的知识能力，为团队助力

你的服务 Agent

Claude CodeCodexCursor托管 Agent 平台

Bot 提示词
Context 技能插件

生产更新检索

远程检索知识发布
  • 知识包
  • 知识网站
  • Git 工作区

同一批真实问题，更准、更快、更省

大型团队和生产 Monorepo 的真实开发者问题，对照纯 LLM Wiki 方案 + 代码库方案测评方法 →

完全解决

LLM Wiki19题Context23题

+16.7 个百分点

平均耗时

LLM Wiki223.2秒Context93.2秒

缩短 58.2%

平均工具调用

LLM Wiki77.5次Context7.8次

减少 89.9%

每题成本

LLM Wiki$2.45Context$0.27

降低 88.9%

01

更准：知识即索引

通过 Indexer 技能，把不同来源、不同技术栈的内容抽取为高压缩的知识索引，并登记证据链写给 Agent 的知识 →

Indexer 按需提取

orders/ · 订单服务

  • api/routes.ts接口与路由
  • service/refund.ts业务规则与调用链
  • auth/policy.ts权限判断
  • config/flags.ts开关与默认值
  • gen/*.pb.ts生成代码
  • styles/样式

Indexer知识：退款流程

压缩率

1/115

来源约 202 MB
584 万行代码 · 1,129 份文档
知识约 1.75 MB
220 篇已审核文章

证据链 →

knowledge / 退款流程

  1. 退款要求订单已支付且未超过 30 天。1

    service/refund.ts:42-58@8f3c2a1

  2. 超过 500 元的退款需主管审批。2

    auth/policy.ts:17-24@8f3c2a1

  3. 退款结果通过 Webhook 通知商户。3

    运维手册 §3v2.4

每段结论都登记出处：文件、行号与版本

02

更快、更省：先读知识，再远程取原文

通过知识快速了解全局，再沿证据链远程抵达原文，精确推理固定版本的搜索与阅读 →

一次问答的两种路径

Agent 沙箱 · 不克隆仓库知识足够，直接作答Source Graph 远程检索证据链知识仓库知识仓库代码仓库 @commit代码仓库 @commit提问提问读知识读知识读原文读原文作答作答

单次查询耗时

本地 rg10–740 ms
本地 git grep9–312 ms
远程检索0.9–2.1 ms

服务端耗时；同机房含网络往返在 100 ms 以内

多版本
同一仓库多个版本并存，按 commit 直接查
结果一致
检出结果与本地搜索相同

KNOWLEDGE MAP

Site map网站地图

Start with a topic, then continue to the pages within each section.从主题进入知识，也可以继续查看栏目中的具体页面。

RESOURCES

Tools and resources工具与资源

LLM Docs

Open docs打开文档

Structured knowledge entry points for agents and models.面向 Agent 和模型的结构化知识入口。

Changelog

View updates查看更新

Review knowledge versions and update history.查看知识内容的版本变化与更新时间。