FAQ
Does a small module need a central knowledge base?
Not necessarily. A central knowledge base suits knowledge that many people share and use remotely: the team registers its repositories, documents and meeting notes once, and its Bot, documentation site and each member's coding Agent all read the same reviewed knowledge.
Content for a small module or a single maintainer can stay in the repository's docs/. Context can register same-repository documents as repository content without copying them into knowledge.
Knowledge is written by AI. How is quality kept, and does it bloat?
People decide the direction; the Agent handles the detail.
- In the review report, a person judges whether the categories, update scope and granularity are right. Nothing is published before its Diff report is approved.
- Every conclusion records its source (file, line and version), and citations are checked before anything is written into knowledge.
- During production and updates the Agent cross-checks sources. It often finds upstream documents that lag behind the code, corrects the knowledge from the code, and can also write the correction back upstream when asked.
- Knowledge compresses sources around questions and keeps only what answers them. When a knowledge base grows, ask the Agent to reorganize its directories and titles.
See Evidence Chains and Index First and Knowledge for Agents.
How is a knowledge package different from an LLM Wiki?
A knowledge package is a container; it does not prescribe how content is written. Context ships default Indexers and writing patterns, and a team can customize its own production skills and the skills a package hands to Agents.
Compared with an LLM Wiki that has only pages, a package also carries an index, evidence chains and an Agent-facing query entry. Iterations of the knowledge base maintain these directories, so nobody syncs them by hand. See Packages and Sites.
The update process looks long. What does a maintainer actually handle?
Only review. Committing the Context knowledge workspace, building the package and publishing the site run quietly by default; the maintainer is interrupted only for the Diff review, a permission problem, or an objection. Once the Diff report is reviewed on the website, the process continues on its own.
常见问题
小模块也需要中心知识库吗？
不一定。中心知识库适合多人共享、需要远程使用的知识：团队把仓库、文档和会议纪要登记一次，团队的 Bot、文档站点和每位成员的编码 Agent 读到的都是同一份经过审核的知识。
小模块或只有一位维护者的内容，放在仓库的 docs/ 里即可。Context 可以把同仓库的文档登记为仓库内容，而不必复制进知识库。
知识由 AI 撰写，质量如何保证？会不会越写越臃肿？
人把握方向，细节交给 Agent。
- 在审核报告中，由人判断分类、更新范围和粒度是否合适。Diff 报告批准之前，任何内容都不会发布。
- 每条结论都记录来源（文件、行号和版本），写入知识之前会先检查引用。
- 在生产和更新过程中，Agent 会交叉核对来源。它经常发现上游文档落后于代码，于是依据代码修正知识；如有需要，也可以把修正回写到上游。
- 知识围绕问题压缩来源，只保留能回答问题的内容。知识库变大后，可以让 Agent 重新整理目录和标题。
知识包和 LLM Wiki 有什么不同？
知识包是一个容器，并不规定内容怎么写。Context 提供默认的 Indexer 和写作模式，团队也可以定制自己的生产技能，以及知识包交给 Agent 使用的技能。
与只有页面的 LLM Wiki 相比，知识包还带有索引、证据链和面向 Agent 的查询入口。这些目录随知识库的每次迭代自动维护，无需任何人手动同步。详见知识包与站点构建。
更新流程看起来很长，维护者实际要处理什么？
只需要审核。提交 Context 知识工作区、构建知识包和发布站点默认都在后台静默完成；只有在 Diff 审核、权限问题或有异议时才会打扰维护者。在网站上审核完 Diff 报告后，流程会自动继续。