A sizable update to the Copilot app on Windows 11, currently shipping to Windows Insiders on Copilot+ PCs, lets you retrieve local files and images by describing what you’re looking for in plain language rather than by fishing for the right filename. The preview, delivered via the Microsoft Store starting with Copilot app version 1.25082.132.0, also remakes the Copilot home screen and tightens the link between the assistant and Vision—the screen-scanning AI that can walk you through an application or your entire desktop.

What the Preview Delivers

Microsoft is rolling out three headline changes:

  • Semantic file and image search inside Copilot. Type something like “find the file with the chicken tostada recipe,” “find my CV,” or “find images of bridges at sunset on my PC” and Copilot returns matches based on meaning, not literal filename tokens. The search leans on a semantic index built from document text and image descriptors.
  • A redesigned Copilot home. The new home pane surfaces recent apps, recent files, and past Copilot conversations, turning the assistant into a launchpad for resuming work. Clicking a recent file attaches it to the chat for summarization or follow-up Q&A.
  • Vision-guided help integration. The “Get guided help” tile can launch a Copilot Vision session that scans the chosen app window—or the entire desktop, with permission—and then walks users through tasks or interprets on-screen content.

The features are gated to Windows 11 Copilot+ hardware with an NPU rated for 40+ trillion operations per second (TOPS) and are initially limited to English (simplified and traditional Chinese are also named in some documents), French, German, Japanese, and Spanish. Supported file formats listed in the Copilot app include .png, .jpeg, .svg, .pdf, .docx, .xlsx, .csv, .json, and .txt.

How Semantic Search Works Under the Hood

Windows has long offered a lexical index that matches filenames and metadata character-by-character. The Copilot app’s semantic search adds a parallel index of vector embeddings—mathematical representations of meaning—generated from document text and visual features extracted from images. When you type a natural-language query, Copilot encodes the query into the same embedding space and performs nearest-neighbor retrieval. The result is a match by concept, not by exact word overlap.

On Copilot+ devices, this inference runs on the local NPU using a small language model Microsoft has tuned for the silicon (the Phi Silica family, described as an on-device SLM). The architecture aims to keep latency low and to reduce cloud round-trips, delivering what Microsoft positions as a privacy-first experience. The preview currently draws results from the Windows Recent folder and other indexed locations by default; it does not spider every file on the machine unless the user explicitly attaches a file into the chat.

The New Home and Vision: A Workflow Hub

Microsoft is reshaping Copilot’s surface from a sidebar chatbot into a workspace. The home page organizes recent activity into three zones: a carousel of recent apps, a list of recent files, and a thread of past conversations. Selecting a recent file fires off an attachment in the current chat, ready for summarization, extraction, or questioning. The “Get guided help” section—stocked with popular Microsoft apps like Excel, Word, and Teams—launches a Vision session that sees the full screen or a chosen window and provides step-by-step assistance. Vision requires an explicit permission grant each time, and Microsoft warns users to close sensitive windows before enabling it.

Privacy and Security: Promises and Open Questions

Copilot’s new reach into local files and the desktop screen raises immediate data-governance concerns. Microsoft’s preview messaging emphasizes a permissions-first model: files are never automatically uploaded; Vision sessions begin only after consent; and on-device processing on Copilot+ PCs keeps data local for compatible queries. The Copilot Settings panel exposes toggles to review what the app can access and retrieve.

Yet several gaps remain. The public preview materials do not fully detail what happens when a query cannot be satisfied on-device—whether a cloud fallback exists, what telemetry is collected, or how long any screen captures are retained. Vision’s ability to scan “everything on your screen” can inadvertently expose password managers, banking portals, or confidential documents if the user isn’t vigilant. Until Microsoft publishes machine-level telemetry and processing-flow documentation, organizations should treat absolute privacy promises as conditional and demand evidence of local-only execution.

Enterprise Checklist for Adoption

IT administrators should treat the combination of semantic indexing, on-device models, and screen-reading AI as a new data plane on managed desktops. Before turning on these features broadly, teams should:

  • Inventory Copilot+ hardware across the fleet to identify which devices can run the local semantic stack.
  • Pilot on non-production machines to verify that Copilot does not automatically upload files or trigger cloud queries unexpectedly.
  • Lock down permissions via Copilot Settings and Windows privacy/search controls, restricting which folders Copilot can index or access.
  • Update data-handling policies to cover explicit attach actions and Vision sessions, and train staff to recognize when consent is given and how to avoid accidental disclosure.
  • Demand documentation from Microsoft or consult forthcoming admin docs for a clear breakdown of telemetry flows, retention periods for Vision captures, and any cloud processing that may occur on non-Copilot+ hardware.

In regulated industries, administrators should plan for selective blocking of Copilot features where off-device processing of certain data classes is prohibited.

Early Wins and Recognized Limitations

The preview delivers clear productivity gains:

  • Faster fuzzy retrieval. Concept-based search cuts the time spent navigating folder trees or guessing filenames.
  • Inline file actions. Attaching a result to a chat and asking “summarize this contract” or “find all mentions of Q3 targets” compresses multiple manual steps into a single flow.
  • Contextual, screen-aware help. Vision can replace manual screenshot-and-paste cycles with a one-click guided session.

But the rollout also exposes several friction points:

  • Hardware fragmentation. The Copilot+ requirement means uneven experiences across an organization’s fleet; behavior may vary by OEM, SoC vendor, or firmware version.
  • Unclear fallback logic. Non-Copilot+ devices may rely on cloud-assisted processing, yet the boundary between local and cloud is not exhaustively documented in preview materials.
  • Consent visibility. One-click attach or a quick screen scan can lead to accidental data exposure if prompts are ignored or glossed over. Administrators must validate that consent UX is unambiguous and reversible.

Practical Tips for Power Users

  • Enable Copilot permission prompts and verify them in Settings. Before clicking a file in the recent pane, double-check that you intend to attach it.
  • Frame queries with enough context to disambiguate results. “Find PDF resume from 2024 named CV” yields better matches than “find my resume” when you have multiple resume versions.
  • Before starting a Vision session, close any window displaying passwords, banking data, or sensitive internal communications. Vision sees what you see.
  • Check update availability by ensuring the Copilot app is on version 1.25082.132.0 or later, and that you are enrolled in a Windows Insider channel with a supported Copilot+ PC.

The Broader Microsoft AI Picture

This test is not an isolated play. Microsoft is methodically building a tiered AI ecosystem in Windows: broadly available Copilot features for standard hardware, and premium, NPU-accelerated capabilities for Copilot+ devices. The semantic search inside Copilot follows the AI-powered Windows Search already previewed in File Explorer, and it sits alongside developer APIs that expose vector indexing for third-party apps. By making Copilot the unifying interface—for search, document analysis, and screen-aware help—Microsoft hopes to embed AI as a default part of daily Windows workflows, not a bolt-on.

The strategy carries operational complexity. Hardware certification, NPU firmware consistency, developer tooling, and transparent data practices must all converge to avoid user distrust. For now, the Copilot app preview tests both the technology and the user-acceptance boundaries that will define Windows AI’s next chapter.

Conclusion

Microsoft’s test of semantic search in the Copilot app is a substantial product evolution that tackles a real productivity bottleneck: finding files when you remember the idea but not the name. On Copilot+ hardware, it promises fast, local inference and a cohesive home that puts recent work and AI assistance side by side. But the features’ ultimate success will hinge on Microsoft’s willingness to publish granular privacy documentation, clarify fallback behaviors on non-Copilot+ machines, and build consent mechanisms that users and enterprises can trust. For now, Insiders with the right hardware can start describing their files rather than hunting for them—and the rest of the Windows ecosystem will be watching to see if the controls keep pace with the capability.