-
Notifications
You must be signed in to change notification settings - Fork 4
Knowledge Bases and Live Watch
← Wiki home / 手册首页 · English · 中文 · Troubleshooting / 故障排查
Use this page when / 适用情况： adding Markdown sources, deciding between full and incremental builds, tuning live-watch quiet time, or checking whether stopped watchers released their workers.
添加 Markdown 来源、选择全量或增量建库、调整热监听安静时间，或确认停止监听后工作进程已经退出。
Each configured knowledge base has its own workspace state and index paths under the active data root. The source directory remains external and is treated as read-only input during normal indexing.
Use the narrowest stable root that contains the material you actually want to retrieve. Selecting a drive root or a large mixed archive increases build time, index size, file-change noise, and the chance of unsupported files.
Use a full build when:
- adding a knowledge base for the first time
- changing parsing or indexing settings that require complete regeneration
- the app reports that vector or lexical index state is missing or incompatible
- a large external file operation made incremental history unreliable
A full build may include parsing, lexical indexing, embeddings, vector writes, final state checks, and source summaries.
Incremental updates process files that changed or were deleted since the last consistent state. They are useful for ordinary writing sessions but are not a substitute for a full rebuild after major configuration changes.
The normal watchdog mode reacts to actual filesystem events. It does not continuously rescan every knowledge base while idle.
The quiet period starts after the latest save. Repeated saves to the same files reset that timer, so editing activity is merged into one batch. The default is suitable for ordinary editors; increase it when an editor or sync tool produces many rapid events.
Generated Logseq version history, backups, recycle trees, and conflict-trash files are ignored where supported to avoid pointless updates.
Incremental parsing, embedding, and vector writes run in disposable worker processes. After a batch finishes, Windows can reclaim their Commit, RAM, and GPU memory. Semantic batches are serialized to avoid loading duplicate models for multiple watched knowledge bases at the same time.
Idle watch still uses a small observer and event bookkeeping footprint, but it should not continuously run semantic models or consume GPU capacity without file changes.
Global stop requests cancellation for every active watcher. A knowledge base can briefly show a stopping state while a scan or vector write reaches a safe cancellation point. Wait for the final stopped state before closing or reconfiguring the source.
If file updates continue after the final stopped message, collect the activity log and see Troubleshooting.
Removing a source from Caelune should remove the app-owned index/state and the source registration. It must not delete the original external notes. Read confirmation text carefully when clearing or removing any data.
每个知识库在当前数据目录下拥有独立的工作区状态与索引路径。原始来源目录位于外部，正常建库时只作为只读输入。
请选择能够覆盖目标资料的最小稳定根目录。直接选择磁盘根目录或大型混合归档，会增加建库时间、索引体积、文件事件噪音和遇到不支持文件的概率。
以下情况适合全量建库：
- 第一次添加知识库
- 修改了需要完整重建的解析或索引设置
- 软件提示字面索引或向量索引缺失、不兼容
- 大规模外部文件操作导致增量历史不再可靠
全量建库可能包含解析、字面索引、向量化、向量写入、最终状态检查和来源摘要。
增量更新只处理从上次一致状态以后发生变化或删除的文件，适合日常写作，但不能代替重大配置变化后的全量重建。
正常 watchdog 模式只响应真实文件系统事件，空闲时不会不断重扫全部知识库。
安静时间从最后一次保存开始计算。同一批文件连续保存会重置计时器，从而把编辑过程合并为一个增量批次。默认值适合普通编辑器；如果编辑器或同步软件会快速制造大量事件，可以适当增加。
软件会尽量忽略 Logseq 自动版本历史、备份、回收目录和冲突垃圾文件，避免无意义更新。
增量解析、向量化和向量写入在一次性子进程中执行。批次完成后，Windows 可以回收对应的 Commit、内存和显存。多个知识库的语义更新会串行处理，避免同时重复加载模型。
空闲监听仍会保留少量文件观察器和事件记录开销，但在没有文件变化时不应持续运行语义模型或占用 GPU。
批量停止会请求取消全部实际监听器。扫描或向量写入需要到达安全取消点，因此知识库可能短暂显示“停止中”。请等待最终“已停止”再关闭程序或重新配置来源。
如果已经出现最终停止消息后仍有更新，请保存活动日志并查看故障排查。
从星野移除来源时，只应删除软件拥有的索引、状态和来源注册，不应删除外部原始笔记。清理或移除数据前仍应仔细阅读确认文字。
Previous / 上一篇： ← Runtime and Models / Runtime 与模型 · Next / 下一篇： Search and Resources / 查询与资源 →
🏠 手册首页 / Wiki Home · ⬇️ 下载最新版 / Latest Release · 🌐 官网 / Website · 💻 GitHub
隐私提醒 / Privacy reminder: 公开日志、截图或配置前，请移除私人路径、知识库名称、查询内容和检索结果。Remove private paths, workspace names, queries, and retrieved text before sharing logs, screenshots, or configuration.