“Find the contract with the renewal clause” — you type it, and Windows stares back with zero results, even though the file is right there in Documents. Semantic file search was the Copilot+ headline feature: search by meaning, not filename. Yours searches by filename only, which is just… regular search.

Semantic file search combines three systems: the Windows Search indexer (catalogs file contents), the NPU (runs the embedding models), and the downloaded AI models (understand your query’s meaning). Break any one and you get silent fallback to keyword search — no error, just dumb results. Diagnosing which layer failed is the whole task.

This guide fixes semantic file search in technician order: eligibility, index health, the NPU and models, search scope, query technique, and the rebuild path. The same layered logic applies across AI PCs — our Postman on ARM guide covers the shared platform stack, and the text-to-image settings guide covers the AI settings layer — while NPU error code 10 is essential reading if your NPU won’t start at all.

Why this happens

  • Not a Copilot+ PC: Semantic search needs a qualifying NPU — anything else gets keyword search, which is correct behavior, not a bug.
  • Index incomplete or scoped wrong: The semantic engine searches the index — unindexed folders and unfinished indexing mean invisible files.
  • AI models not downloaded: The embedding models are separate downloads. Without them, queries never get “understood.”
  • NPU driver broken: A flagged or missing NPU driver disables the semantic layer silently.
  • Searching unindexed content types: Semantic understanding covers documents, photos, and settings — it won’t semantically search inside exotic file types.
  • Query technique: Overly vague queries (“stuff”) or queries in unsupported languages defeat even a healthy engine.
  • Corrupted index: A damaged index breaks both classic and semantic search identically.

How to fix it

Work through these in order — diagnose first, then act.

Fix 1: Confirm eligibility and test the layers separately

Isolate which layer is broken before fixing anything: hardware, index, or AI models. Each has a different fix, and guessing wastes time.

  1. Confirm Copilot+ in Settings > System > About and a clean NPU in Device Manager > Neural processors.
  2. Test classic search first: search an exact filename you know exists — if even that fails, the index is broken (skip to Fix 6).
  3. Test semantic search: query a concept (“invoice from last quarter”) for a file whose name doesn’t contain those words.
  4. Filename works + concept fails = AI layer problem (Fix 3). Both fail = index problem (Fix 2/6).
  5. If hardware doesn’t qualify, semantic search is unavailable — but index fixes still improve classic search.

Fix 2: Fix index scope and completion

The semantic engine can only understand indexed files. Scope and completion issues are the most common real cause of “semantic search finds nothing.”

  1. Open Settings > Privacy & security > Searching Windows — confirm Enhanced mode (not Classic).
  2. Review Excluded folders — your Documents, Pictures, and working folders must not be excluded.
  3. Check indexing status and item count — plug into AC power and let a partial index finish (hours for large drives).
  4. Verify your file types are indexed: Advanced indexing options > File Types — check your document extensions.
  5. Don’t test semantic queries until the index is complete — the engine can’t see what isn’t indexed.

Fix 3: Restore the NPU driver and AI models

With a healthy index but no semantic understanding, the AI layer is missing. NPU driver plus model downloads, in that order.

  1. Check Device Manager > Neural processors — reinstall the NPU driver from your laptop maker if flagged or missing.
  2. Run Settings > Windows Update with optional updates — semantic models arrive as AI components.
  3. Use unmetered Wi-Fi with free disk space; leave the PC idle, plugged in, 1-2 hours.
  4. Confirm the models landed: semantic queries should now return concept matches, not just filename matches.
  5. If the NPU shows error code 10 or won’t start, fix that first — a dead NPU means no semantic layer, period.

Fix 4: Tune search scope and permissions

Right layers, wrong scope still hides files. Permissions, cloud settings, and per-location indexing all filter semantic results.

  1. Re-verify Excluded folders — remove exclusions for anywhere you keep searchable files.
  2. Check Search permissions in Privacy & security — cloud content and history toggles affect what’s searchable.
  3. For OneDrive/SharePoint files, confirm sync status — online-only files with no local presence may not index fully.
  4. Ensure you’re signed into your Microsoft account for cloud-backed results.
  5. Test the same semantic query in File Explorer vs. Start — if one works, the other’s surface has the bug.

Fix 5: Query like the engine expects

A healthy engine still fails on bad queries. Semantic search understands natural descriptions — but it needs something to work with.

  1. Describe content, not filenames: “presentation about Q3 sales” beats “Q3 deck final FINAL v2.”
  2. Include a content hint: “email about the server outage” works better than “server stuff.”
  3. Use the document’s language — the models are strongest in English; other languages need matching language packs.
  4. Add a time hint when you remember one: “contract from last spring” narrows the semantic space enormously.
  5. If a query fails, try two or three phrasings before concluding the engine is broken — query variance is real.

Fix 6: Rebuild the index and verify end-to-end

When the index itself is corrupt, only a rebuild fixes it. Then verify both layers together — index plus AI — because semantic search needs both simultaneously.

  1. Go to Searching Windows > Advanced indexing options > Advanced > Rebuild and confirm.
  2. Plug in, leave idle — full rebuilds take 2-6 hours on large drives.
  3. After rebuilding, re-verify AI models are still downloaded (Fix 3) — rebuilds don’t touch models, but confirm anyway.
  4. Run the full test battery: exact filename (classic layer), then concept query (semantic layer).
  5. If semantic search still fails after a clean rebuild with healthy NPU and models, check the Microsoft Windows support hub for build-specific known issues.

Closing

Semantic file search is a three-layer system — indexer, NPU, AI models — that fails silently to keyword search when any layer breaks. Test the layers separately to find the broken one, fix index scope and completion, restore the NPU driver and models, tune your scopes, query descriptively, and rebuild the index when it’s corrupt. When all three layers are healthy, search finally finds files by what they mean.

Frequently asked questions

What is semantic file search?
It’s Copilot+ PC search that understands meaning — “budget spreadsheet from March” finds the file without the exact name. It uses on-device NPU models over the Windows Search index.
Why does search only match filenames?
Filename-only matching means the semantic layer isn’t active — the NPU driver, AI models, or indexing is incomplete. The search box falls back to classic mode silently.
Does semantic search work in File Explorer?
Yes — File Explorer’s search box uses the same semantic engine on Copilot+ PCs once indexing and models are ready. Both Start menu and Explorer share the backend.
Which files does semantic search understand?
Documents (text content), photos (visual content on Copilot+ PCs), and settings. It works best on indexed text documents and images with the AI models downloaded.