最难找的文件，往往恰好是当初命名最随意的那一个。
截图叫 Screenshot 2026-09-28…，照片还是 IMG_2107.jpg，一份有用的笔记混在几十个 Markdown 文件里。过了几周，我还记得里面写了什么、画面是什么样，却想不起文件叫什么。
这也是我开发 AnySearch 这个 macOS 原生搜索工具时，一直在解决的问题。最初的目标是更快地按文件名查找；后来更有意思的问题变成了：当文件名不再是线索时，搜索应该依赖什么？
这篇文章梳理一下这条演进路线：基础文件搜索、3.0 的文件全景与自然语言条件解析、为 3.2.0 实现的本地 AI 内容检索，以及为 3.2.1 准备的应用搜索。
先说明版本边界：应用搜索属于后续开发工作，本文不会把所有能力都当作已经在公开商店版本中上线。具体可用功能以当前版本更新说明为准。
一、先把精确搜索做好，不是所有问题都需要 AI
如果我知道目标是一张最近修改过的 PDF 发票，就不需要模型猜测，只需要这些条件能够准确地组合起来。
AnySearch 使用自己的本地 SQLite 索引，不依赖 Spotlight 的索引。文件系统变化通过 FSEvents 更新到索引里，因此搜索是在已有数据库上查询，而不是每次重新遍历目录。
普通搜索支持包含、前缀、后缀、精确匹配，也可以组合大小写、扩展名、大小、日期、目录和文件类别等条件。例如：
invoice ext:pdf modified:7d
report ext:pdf NOT draft size:>1mb
kind:image in:~/Downloads/
三行分别表达：
- 最近 7 天修改过、名称包含 invoice 的 PDF；
- 名称包含 report、排除 draft、大于 1 MB 的 PDF；
- 下载目录中的图片。
还可以使用 OR、引号短语、通配符和创建日期筛选，逐步缩小结果范围。
这里，可视化筛选和查询语法同样重要。日期、目录、扩展名应该是用户能看见、能修改的条件，而不是藏在某个排序算法里的假设。
二、“自然语言搜索”其实可能是两件不同的事
有个区别很容易被一张 AI 功能清单掩盖：把一句话转换成条件，不等于按内容含义检索。
3.0 中的自然语言查询，例如“查找上周大于 1 MB 的发票”，由 Mac 上的确定性规则处理。程序识别支持的时间、大小、类别和目录表达，把它们转换成结构化过滤条件。
这让已有的搜索引擎更容易使用，但不意味着引擎已经能理解照片里的画面。
解析后的条件会显示为标签。如果严格条件没有结果，程序尝试放宽日期、目录或类别时，也会展示采用的策略，而不是偷偷换一个问题。
我希望保留这个区别：已知条件足够明确，就用可解释的精确筛选；真正缺少的是文件名线索时，再使用语义检索。
三、只记得内容，也应该能找到原文件
AI 演示里，一个直观的例子是 IMG_2107.jpg。
输入“戴墨镜的猫”，描述的是照片内容，而不是文件名。在实际录制的演示中，程序找到了这张照片，并在右侧预览。
动图 1：用画面描述找照片，不依赖文件名。
演示中还有“支付失败的截图”“吃火锅的发票”“去杭州旅游的笔记”等查询。这些例子的共同点是：人对内容的记忆，通常比系统自动生成的文件名更有辨识度。
当然，这些是小规模准备素材上的功能演示，不是对任意个人文件库的检索准确率保证。
下面几段录屏分别展示不同类型的内容线索。
按场景找截图
动图 2：不需要记住截图的日期和自动生成的名称。
按用途找发票
动图 3：线索是发票对应的事情，而不是一长串文件编号。
不只找视频，还能定位到匹配画面
动图 4：查看匹配的视频时刻，也能对比同一主题下的图片。
按笔记内容找文档
动图 5：找到原文档后，直接在预览中确认是不是要找的那一份。
为 3.2.0 实现的本地 AI 搜索，将这个思路扩展到照片、截图、PDF 页面、文档段落、EPUB 书籍，以及音视频片段。可选的代码搜索使用 code: 前缀。对支持的结果，程序会展示匹配的页码、段落或时刻，而不是只扔给用户一个大文件。
底层做的是检索，不是让聊天机器人生成回答：
- 本地 embedding 模型把内容和查询编码成向量；
- 搜索引擎结合语义信号，以及文件名、提取文本、OCR 文字上的全文匹配；
- 返回原始文件，并尽量定位到命中的内容位置。
这两类信号需要互补。精确的标识符、少见的词，不应该因为另一份文件“语义上有点像”就被淹没；一张照片也不应该先手写好说明文字，才具备被搜索的资格。
四、“本地运行”需要讲清楚具体含义
AI 搜索在 Apple silicon Mac 上，通过 llama.cpp 运行 EmbeddingGemma 2。第一次使用需要下载约 865 MB 的模型文件，不是说一个体积很小的安装包里已经悄悄包含了所有模型。
模型安装完成后，AI 索引和搜索可以离线运行：
- 查询在 Mac 上编码；
- 提取的内容和向量保存在本地索引中；
- 只索引用户选择的目录；
- 不需要注册 AI 账户，也不需要配置 API Key。
模型下载站点会收到提供文件下载所必需的普通连接信息，但不会收到正在索引的文件内容或之后输入的搜索查询。
本地运行也不是没有成本：初次建索引需要时间，模型冷启动比热搜索慢，模型和索引占用磁盘，计算也会消耗内存和算力。
普通文件名搜索继续支持 Intel Mac；本地 AI 搜索要求 Apple silicon。音频检索也还有局限，当前实现中，中文口语音频的语义检索效果仍弱于文档和照片。
这些边界，比一句“什么都能搜到”更有参考价值。
五、索引还在更新时，搜索也应该能正常用
一个搜索框在静态演示中很快，并不代表它在真实建索引过程中也好用。
最近的一部分引擎工作，就在处理这些不显眼但影响体验的问题：
- 用户继续输入后，取消已经过时的查询；
- 新结果返回前保留现有列表，避免反复清空；
- 用户切换选中项后，不让旧预览覆盖新预览；
- 后台刷新向量时，不让交互查询排在刷新任务后面等待。
向量刷新现在会在后台构建独立的只读数据库快照。交互搜索继续使用当前快照，新快照构建完成后再切换。
一次本地回归测试使用了 18,668 个向量，8 个查询各重复 10 次。在后台刷新期间，检索延迟中位数从约 128 ms 降到约 10 ms。80 次前后对比中，返回文件与匹配片段的排序保持一致。
这里必须说明测量边界：查询向量已经提前生成，测的是检索和排队延迟，不包含模型编码、冷启动、输入防抖和界面渲染。不能据此宣称“任何 Mac 上的完整 AI 搜索都只需 10 ms”。
真正想达到的效果很朴素：修改查询、切换结果时，不要像是在和索引器抢使用权。
六、有时先看清文件分布，比直接搜索更有用
搜索默认我知道自己要问什么。File Landscape（文件全景）解决的是还不知道该从哪里找的问题。
3.0 引入的文件全景，会按类别、扩展名、大小、目录和最近活动汇总已索引文件，也提供大文件和可能重复项入口。点击统计项，可以回到主搜索并带上对应条件。
动图 6：从文件类型、空间占用切换到活动趋势，画面已裁掉个人目录区域。
它不是一个看完就结束的独立仪表盘，而是开始搜索的另一种方式。
这里也有两个限制：
- 统计对象是已授权、已索引的集合，不是对整个系统盘的完整扫描；
- “可能重复”基于文件名和大小提供候选，不代表已经验证文件内容完全一致。
对实际清理任务来说，这往往足够作为起点：先看空间主要被哪类文件占用，再检查大文件，最后决定如何处理原文件。
七、应用也有自己的“名字记不全”问题
后续为 3.2.1 准备的能力，是应用搜索。
人们记住一个应用的方式，可能是完整名称、简称、旧名称，也可能只是几个字母。匹配器因此支持名称、前缀、子串、有序缩写、有限的拼写容错，以及中文名称、拼音和拼音首字母。
应用已安装的前提下，例如：
应用会出现在普通搜索和快捷面板中，与文件一起呈现。但输入 ext:pdf 这样的明确文件查询时，不应该混入无关应用。
这项工作的范围，比做一个命令启动器或插件平台小得多。它读取应用包元数据和本地化名称，不需要读取应用内部的用户业务数据。非标准安装目录需要单独授权，而不是绕过 macOS 权限。
八、找到结果，还不代表搜索任务完成了
无论是哪种搜索模式，我最终想保留的流程都是：
输入 → 辨认结果 → 确认内容 → 继续工作。
因此，文件名旁边要能看见所在目录，键盘选择要稳定，Quick Look 要方便调用，打开文件、在 Finder 中显示、复制路径和拖入其他应用这些动作，都应该容易触达。
如果正确结果已经出现，却还需要花一分钟确认它到底是什么，那么搜索问题并没有真正结束。
AnySearch 的方向，不是在每次查询前面都加一个 AI，而是根据用户还记得什么，使用合适的检索方式：
- 记得明确条件，就用文件过滤；
- 只记得内容，就按内容检索；
- 不清楚文件分布，就先看全景；
- 只记得应用缩写，也能找到对应应用。
如果想体验，应用在 Mac App Store 中的名称是 AnySearch Pro。后续 AI 和应用搜索能力的可用情况，请查看当前版本更新说明。
本文由 AnySearch 开发者撰写。截图和动图来自实际演示及开发版本，用于说明功能行为，不代表通用检索准确率评测。动图为录屏节选与压缩版本，不用于衡量端到端搜索耗时。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The most difficult document to find often coincided with the most randomly named one.
The screenshot is Screenshot 2026-09-28... and the picture is also IMG 2107.jpg, a useful note mixed with dozens of Markdown files. After a few weeks, I remember what it says, what it looks like, but I can't remember what it's called.
And that's what I've been working on when I developed AnySearch, this macos primary search tool. The initial objective was to find the file name more quickly; later the more interesting question became: What should a search rely on when the filename is no longer a clue?
This article combs this evolutionary path: basic document search, document panorama and natural language interpretation, local AI content retrieval for 3.2.0, and application search for 3.2.1.
First state the boundaries of the version: the application search is a follow-up development, and this paper will not consider all capabilities as already online in the open store version. Specific available functionality is based on the current version of the update.
First, let's get a precise search done. Not all questions need AI.
If I knew that the target was a recently modified PDF invoice, there was no need for model speculation, just those conditions that could be accurately combined.
AnySearch uses its own local SQLite index, which does not depend on Spotlight. The file system changes are updated to the index through FSEevents, so the search is queryed on an existing database rather than revisiting the directory every time.
Normal search support contains, prefixes, suffixes, precise matching, and can also combine conditions such as case, extension, size, date, directory and file category. For example:
invoice ext:pdf modified:7d
report ext:pdf NOT draft size:>1mb
kind:image in:~/Downloads/
Three lines say:
- The last seven days of the revision of the PDF, whose name contains the reference;
- The name contains the report, excludes the PDF, which is larger than 1 MB;
- Download pictures from the directory.
The results can also be progressively narrowed using OR, quote phrases, wildcards and creation dates.
Here, visual screening and search syntax are equally important. Date, directory, extension should be a condition that the user can see and modify, rather than an assumption hidden in a sorting algorithm.
“Natural Language Search” may actually be two different things.
There is a distinction that can be easily obscured by an AI functional list: converting a sentence into a condition does not amount to searching by content.
Natural language queries in 3.0, e.g., “Find invoices greater than 1 MB last week”, are handled by the certainty rule on Mac. The program identifies the time, size, category and directory expressions supported and converts them into structured filter conditions.
This makes existing search engines easier to use, but does not mean that the engines already understand the picture.
The condition will be displayed as a label after resolution. If the strict conditions are not met, the programme, when attempting to relax the dates, directories or categories, will also demonstrate the strategy used, rather than swipe a problem.
I would like to retain this distinction: the conditions are known to be sufficiently clear to allow for precise and interpretable screening; what is really missing is the use of semantic searches.
III. Just remember the contents, and you should find the original document.
In the AI demonstration, an intuitive example is IMG 210.jpg.
Enter " Cat in sunglasses ", describing the photo content, not the file name. In the actual recorded presentation, the program found the photo and previewed it on the right.
Anticipated #1: Image description for photo search, not by filename.
The presentation also included a survey of “failed payment screenshots” “Invoices to eat hot pots” “Notes to travel to Hangzhou”. The common denominator of these examples is that the memory of content is generally more visible than the file name generated automatically by the system.
These are, of course, performance demonstrations on small-scale preparation of asset, and are not guarantees of the accuracy of any individual file library.
The following paragraphs show different types of content threads.
We'll take the scene.
Figure 2: There is no need to remember the date of the screenshot and the name automatically generated.
Invoices by purpose
Figure 3: Threads are the subject of invoices rather than a long series of document numbers.
It's not just a video. It's a match.
Anticipated 4: Viewing a match of video moments can also compare pictures under the same theme.
Search for documents by taking notes
Antogram 5: Once the original document has been found, confirm directly in the preview whether it is the one you want.
Local AI search achieved for 3.2.0 extends this idea to photos, screenshots, PDF pages, document paragraphs, EPA books, and audio and video clips. Optional code search using code: prefix. For the supported result, the program displays a matching page number, paragraph or moment, rather than throwing only one large file to the user.
The bottom line is to retrieve, not to allow chat robots to generate answers:
- Local embedding model encodes content and queries into vectors;
- The search engine combines semantic signals and full text matching of file names, extracting text, OCR text;
- Return the original file and locate as much as possible to the target content.
These two types of signals need to be complementary. Accurate identifiers, rare words, should not be flooded because another document “symmetrically resembles”; nor should a photograph be able to be searched before it was written in writing.
The “local operation” needs to be specified.
AI search on Apple silicon Mac to run Embedding Gemma 2. The first use of a model file requiring downloading of about 865 MB is not to say that all models are quietly included in a small installation package.
Once the model is installed, the AI index and search can run offline:
- Encoding on Mac for queries;
- The contents and vectors extracted are kept in local indexes;
- Index only the directories selected by the users;
- No registration of AI accounts or configuration of API Key.
Model download sites receive normal connection information necessary to provide file downloads, but do not receive search queries that are being indexed or entered later.
Local operation is also not cost-neutral: initial indexing takes time, model cold start is slower than thermal search, models and indexes occupy disks, and calculations consume memory and computing.
Normal filename search continues to support Intel Mac; local AI search request Apple silicon. Audio retrieval is also limited and the semantic search of Chinese-language voice is still less effective than that of documents and photographs.
These borders are more of a reference than a phrase “can find anything”.
V. When the index is updated, the search should also be operational.
A search box is fast in static demonstrations and does not mean that it is useful in real indexing.
Part of the recent engine work is addressing these issues, which are not visible but affect experiences:
- Cancel outdated queries after users continue to enter them;
- Keep existing lists before the new results return and avoid repeated clean-up;
- Not to allow old previews to overwrite new previews after users switch the selection;
- When the backstage is refreshed, no interactive queries will wait behind the refreshing task.
Vector refreshing will now build independent read-only database snapshots in the backstage. The interactive search continues with the current snapshot, which will be switched after the new snapshot is constructed.
A local regression test used 18,668 vectors and 8 queries repeated 10 times each. The median search delay dropped from about 128 ms to about 10 ms during the background refreshing period. Returns the file in the 80-to-exceed order of the matching clip.
Here, it is necessary to describe the measurement boundary: the search vector has been generated in advance and the search and queue delays are measured, excluding model code, cold startup, entering tremors and interface rendering. This cannot be said to mean "any full AI search on Mac will only require 10 ms".
The real effect is simple: when changing queries, switching results, do not act like you're taking access to an indexer.
VI. Sometimes seeing the distribution of documents first is more useful than searching directly
I know what to ask myself. File Landscape (document panorama) addresses a problem that is not yet known where to find it.
3.0 The document panorama introduced will summarize the indexed documents by category, extension, size, directory and recent activity, and will also provide large documents and potential duplicate entry points. Click on the statistics entry to return to the main search and take the corresponding conditions.
Figure 6: From file type, space occupancy to activity trend, the image has been reduced.
It is not an independent dashboard that ends after reading, but an alternative way to start searching.
There are two limitations:
- The statistical target is an authorized, indexed collection, not a complete scan of the entire system;
- The “possible repetition” is based on the availability of candidates based on the name and size of the document and does not mean that the content of the document has been verified as fully consistent.
This is often sufficient as a starting point for actual clearance tasks: First looking at which type of document the space is primarily occupied, then checking the large document and finally deciding how to handle the original document.
Application also has its own “name insufficiency” problem
The subsequent capability of 3.2.1 readiness is to apply search.
One way to remember an application may be a full name, abbreviation, an old name or just a few letters. Matchers therefore support names, prefixes, substrings, orderly abbreviations, limited spelling errors, and Chinese names, spellings and spelling initials.
Apply installed conditions such as:
Applications are presented with files in the normal search and shortcut panel. But when entering a specific file query like ext:pdf, you should not mix into unrelated applications.
The scope of this work is much smaller than making a command starter or plug-in platform. It reads application package metadata and localized names and does not need to read in-house user business data. Non-standard installation catalogues require separate authorization, rather than bypassing MacOS permission.
Eight, finding the results doesn't mean the search is complete.
Whatever the search mode, I'd like to keep the process:
Enter the result of the identification confirmation content to continue work.
So, if you can see the directory next to the filename, if the keyboard selection is stable, if Quick Look is easy to call, if you want to open the file, if you want to show it in Finder, if you want to copy the path, and if you want to drag it to another application, you should be easy to reach.
If the correct result has appeared, but it takes one minute to ascertain what it is, then the search problem is not really over.
The direction of AnySearch is not to add an AI to every query, but to use the appropriate search method depending on what the user remembers:
- filtration of documents with clear terms;
- Retrieving content by content only;
- The distribution of documents is not clear, so we look at the landscape first;
- Just remember to apply the abbreviations and find the corresponding application.
If you want to experience it, the name used in Mac App Store is AnySearch Pro. For information on the availability of follow-up AI and applications of search capabilities, see the current version for an update.
This is by AnySearch Developer. Interceptions and animations are derived from actual presentations and developed versions that are used to describe functional behaviour and do not represent universal search accuracy assessments. The animation is a selected and compressed version of the tape, which does not measure the time spent searching from end to end.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.