Microsoft has begun testing a new breed of file search inside its Copilot app on Windows 11—one that lets you describe what you want to find in plain English rather than typing exact filenames. The update, rolling out gradually to Windows Insiders as Copilot app version 1.25082.132.0, introduces conversational, semantic search for local files and images, along with a redesigned home page that surfaces recent files and apps. For now, the full on‑device experience is exclusive to Copilot+ PCs, the company’s latest hardware tier built around a Neural Processing Unit (NPU) delivering at least 40 trillion operations per second (TOPS).

The feature represents a fundamental shift in how Windows handles search, moving from literal string matching to intent‑aware retrieval powered by vector embeddings. Instead of remembering that your résumé is saved as “CV_final_v3.docx,” you can ask Copilot to “find my CV” and get a ranked list of matches. The underlying technology also searches images by visual descriptors—queries like “photos of sunsets with bridges” work even if the files have generic names like IMG_4021.jpg.

What Microsoft shipped in the preview

Two headline changes appear together in this flight. First, the natural‑language search itself, which is accessible directly from the Copilot app. Results appear inline, and clicking a file attaches it to a chat for summarization, Q&A, or extraction. Second, a new Copilot home screen that functions as an AI‑powered dashboard: it shows a user’s most recent apps, recent files, and conversation history, alongside a “get guided help” panel that can launch Copilot Vision sessions for contextual, step‑by‑step assistance on the desktop.

Vision integration is especially notable. When you select an app from the guided‑help area, Copilot can, with your permission, scan the window or the entire desktop and walk you through tasks. The redesigned home ties this capability directly to fast‑access workflows, making it easier to find a file and immediately act on it—for instance, open a spreadsheet, then have Copilot analyze its contents—all within a single chat thread.

In the preview, file upload and preview support is limited to common formats: .png, .jpeg, .svg, .pdf, .docx, .xlsx, .csv, .json, and .txt. Natural‑language comprehension is optimized for English, Simplified Chinese, French, German, Japanese, and Spanish. Microsoft says broader format and language support will follow in later builds.

How semantic search works under the hood

The new system builds on the classic Windows Search index by adding a second, semantic layer. Traditional indexing still handles filenames, metadata, and literal full‑text matches. The semantic index stores vector embeddings of file content, generated by a local AI model running on the device’s NPU when available. Image files are described through computer‑vision object labels and visual descriptors, enabling text‑based search against photos.

When you type a query, Copilot converts it into an embedding and performs a nearest‑neighbor lookup against the semantic index. This surfaces items whose meaning aligns with your intent, even if none of the keywords overlap. Microsoft says the NPU routing reduces latency and keeps routine queries offline, a design choice that also limits the privacy surface: embeddings and index artifacts remain local, never transmitted to the cloud unless you explicitly attach a file to a chat session.

File scope in this preview is deliberately conservative. Results draw primarily from the Windows “Recent” folder and any locations configured under Settings > Privacy & security > Searching Windows. Content sitting in unindexed folders or cloud‑only OneDrive storage (integration is planned for future flights) won’t appear. This reduces indexing overhead and limits accidental exposure while early adopters test the feature.

Privacy, permissions, and the on‑device promise

Privacy messaging is front and center. Microsoft states that only compatible, recently accessed files are surfaced, and nothing is shared with Copilot until a user explicitly clicks to attach. The Copilot Settings page exposes toggle controls for app access, and the company highlights local NPU inference as a privacy boon—queries and result generation can happen entirely on the chip, with no network call.

These protections echo the approach of other Copilot+ AI features such as Recall (preview), which takes encrypted screen snapshots and processes them locally for timeline search. On Copilot+ PCs with Windows Hello Enhanced Sign‑in, Recall stores snapshots on the hard drive, encrypted, and uses similar on‑device processing to support “Click to Do” actions and timeline navigation. The parallel is instructive: Microsoft is systematically building an AI layer that relies on the NPU to keep sensitive data off the network.

Nevertheless, security‑conscious users and IT administrators should examine the details. Microsoft’s blog posts note that embeddings and semantic indexes are stored locally, but they do not yet detail encryption at rest or binding to platform security modules like TPM or Pluton. For enterprises, that leaves open questions about forensics and compliance: if an auditor asks where indexing artifacts reside and who can access them, the answer remains partially obscured. The company says more documentation will arrive as the feature matures.

Another friction point is the attachment model. While the “click to attach” step is an explicit user action, human error still poses risk. An employee might accidentally attach a sensitive contract to a Copilot chat while trying to find it. Organizations should therefore plan to layer data‑loss prevention (DLP) and conditional access controls around Copilot usage.

Hardware gating and what it means for users

Copilot+ certification is the gatekeeper for the best experience. Devices must include an NPU with 40+ TOPS, Pluton security integration, and a minimum 256 GB drive with 50 GB free for AI features like Recall. Early Copilot+ PCs shipped with Qualcomm’s Snapdragon X series; AMD and Intel platforms with equivalent NPUs followed later. Outside this hardware tier, Windows 11 users may still get a degraded version of semantic search that relies on the cloud or a less responsive CPU path, but Microsoft hasn’t committed to a feature‑parity timeline.

This staggered rollout is pragmatic—NPU‑based inference truly delivers low latency for real‑time search—but it risks fragmenting the Windows experience. A user on a two‑year‑old business laptop will not see the same instant, offline semantic results as their colleague with a newly purchased Copilot+ device. For organizations planning hardware refreshes, the feature becomes another incentive to move to NPU‑enabled silicon.

Enterprise and IT recommendations

For IT teams, the preview requires a measured pilot. Deploy a small group of Copilot+ devices, then document what the Copilot app touches. Key steps:

  • Review and tighten indexed locations via Settings > Privacy & security > Searching Windows. The default inclusion of the Recent folder can inadvertently expose temporary files or drafts.
  • Validate endpoint logging: determine whether file names, embedding metadata, or query strings appear in telemetry streams. These details will matter for internal audits and privacy impact assessments.
  • Implement DLP policies that block or warn on file attachments to AI chat interfaces, supplementing the user‑consent model with a guardrail.
  • Monitor backup and disaster recovery systems. Semantic index files could inflate backup snapshots if not excluded.

For consumers, the recommendations are simpler: check Copilot Settings permissions, limit indexing to non‑sensitive folders during the preview, and treat the feature as experimental until production documentation lands.

Accuracy, limits, and the UX trade‑offs

Early builds of semantic search aren’t magic. Their accuracy depends on language, file complexity, and how well the embedded model generalizes to niche document types. The six‑language optimization means non‑supported languages will fall back to far less accurate matching. Similarly, files outside the indexed scope—a research paper stored in an external hard drive, for example—won’t appear. False positives are another trade‑off: a similarity‑based system might surface a document containing personal health information because its vector is adjacent to the query, even if the document isn’t the intended target. This isn’t a data leak per se, but it’s an unintended exposure that underscores the need for careful scoping.

The Vision and guided‑help flows, while powerful, introduce new UI challenges. Sharing a desktop view with Copilot is a potent support tool, but it demands trustworthy consent dialogs and clear visual indicators that a session is active. Microsoft’s design team will need to iterate on these signals to prevent accidental over‑sharing.

How to try it now

Insiders can test the feature today, provided they have a Copilot+ PC. Steps:
1. Enroll the device in a supported Windows Insider channel.
2. Update the Copilot app to version 1.25082.132.0 or later via the Microsoft Store (the rollout is staged, so not every Insider will see it immediately).
3. Open Copilot Settings > Permissions to review access controls.
4. Optionally, adjust the Windows Search index scope under Settings > Privacy & security > Searching Windows to exclude sensitive folders.

Windows discovery gets a neural backbone

Microsoft’s semantic search preview is more than a quality‑of‑life upgrade; it signals a strategic shift away from keyword‑based retrieval toward an AI‑mediated discovery layer. By pairing a local vector index with a conversational prompt interface, the company is betting that users will abandon traditional Explorer searches and third‑party tools in favor of a single, intent‑aware pane. The bet is hedged on Copilot+ hardware: only devices with robust NPUs can deliver the privacy and latency benefits needed to make the experience feel native.

Competitors are moving in the same direction. Screen‑aware assistants from Apple, Google, and Samsung are already turning natural language queries and spatial awareness into table‑stakes features. Microsoft’s advantage—an operating system‑level index that feeds into a persistent, side‑panel‑style assistant—could differentiate Copilot+ PCs if the execution holds.

For now, the preview is a promising but unfinished product. The core mechanics work, and the productivity gains for tasks like finding old receipts or family photos are evident. But unanswered questions about index security, cloud fallbacks, and enterprise governance mean the feature should be approached with cautious optimism. Those who test it on eligible hardware will get an early taste of a Windows where search understands what you mean, not just what you type.