[CmdPal] Phase 7c: early-frame relevance (withhold weak fuzzy tail on ultra-short queries) by michaeljolley · Pull Request #49249 · microsoft/PowerToys

Michael Jolley (michaeljolley) · 2026-07-10T05:35:51Z

Note

Last in a series rearchitecting the MainListPage search and ranking logic. The design detail lives in the initial PR #49189.

What's going on

Type one or two characters into Command Palette and you'll watch it show you obviously wrong stuff for a beat before the right results land. Everybody's seen it. This is the fix.

Here's the why. A single character query fuzzy matches a huge slice of the app catalog, and all of it lands in one Fuzzy tier. On a one character match the lexical quality is basically zero, so FrecencyScale 1 beats LexicalScale 10 times almost nothing. Frecency ends up deciding the order inside that giant tier, and your most used app floats to the top on a weak mid word match. There's a new test that reproduces exactly that: for a query that only matches mid word, every result lands in the Fuzzy tier, and seeding frecency on one app floats it to rank one.

Global fallbacks aren't the culprit, which surprised me. They always classify at FallbackFloor and can never leapfrog commands or apps. The rest of the reshuffle is a late async app batch folding in and floating a weak fuzzy app, same root cause.

The plan

In GetSearchViewItems, when the query is one or two characters, hold back the low confidence app tail and show only apps at tier AcronymWordBoundary or better. That's word boundary, prefix, exact, and alias. Commands and fallbacks render exactly as they did.

The scored arrays are already sorted descending and the tier sits in the packed score's high bits, so the kept set is a contiguous prefix. That makes this one forward scan and an ArraySegment view handed to the result factory. Nothing gets re scored, nothing gets reordered, and no ranker constant, tier, or frecency value moves.

The same gate runs on a late async app batch, which is what stabilizes the short query fold in.

At three characters or more the gate is a no op that hands back the same array instance, so the settled order is byte for byte what 7b produced.

EarlyFrameRelevanceTests covers the frecency floats a weak match evidence, the gate withholding the fuzzy tail, the gate keeping confident tiers, three plus character pass through against the same reference, the length boundary, the null and empty and zero length no ops, and the prefix length helper. The 7b equivalence tests stay green and unchanged.

Notes for reviewers

There's a more direct fix I didn't ship. Damping frecency or re gating tiers for short queries inside the ranker would attack the root cause instead of hiding it at render time. It would also change the settled order and break the 7b equivalence test, so it isn't something I want to sneak in under a perf PR. Worth a real conversation on its own.

This is a render path gate, and I want to be honest that it's a gate. The ranker still thinks your most frecent app is the best answer for c. We just stop drawing that opinion until the query says something.

Michael Jolley (michaeljolley) · 2026-08-03T21:11:22Z

/azp run

azure-pipelines · 2026-08-03T21:11:29Z

Azure Pipelines: 1 pipeline(s) were filtered out due to trigger conditions.

…red ranker (#49189) >[!WARNING] > This PR is one in a series of PRs focused on rearchitecting the search/scoring logic of the `MainListPage`. An explanation of the entire search/scoring logic can be found below. > > **This PR should not be merged until PR #49190 is merged into it.** >[!NOTE] > To test the final result, run the branch associated with PR #49249. This stack rebuilds how Command Palette ranks and displays results on its main page. Strong text matches now consistently appear above weaker ones. Usage history and provider preferences can improve ordering between similarly relevant results, but they cannot push a poor match above an obvious one. The stack also makes search feel faster. Results appear without waiting for slower providers, app scoring runs more efficiently, and weak matches are hidden while the user has typed only one or two characters. Automated tests protect the new behavior, while privacy conscious telemetry measures performance and relevance without recording searches. ## Pull requests 1. [#49189](#49189) introduces the new ranking foundation. Results are grouped by match strength, ensuring exact names, prefixes, and acronyms rank above loose fuzzy matches. 2. [#49190](#49190) improves how Command Palette learns from command usage. Recent and frequently used commands receive a sensible boost, and that history now persists across restarts. 3. [#49191](#49191) lets users give each provider a Lower, Normal, or Higher search preference. This preference helps resolve close matches without overriding result relevance. 4. [#49194](#49194) makes the first set of results appear sooner. Commands and apps are shown immediately, while slower fallback results are added when they become available. 5. [#49195](#49195) adds a comprehensive relevance test suite. It verifies that common searches return the expected results and protects ranking quality from future regressions. 6. [#49197](#49197) adds privacy conscious search telemetry. It measures result counts, response time, and which result position was selected without recording search text, result names, paths, or other user content. 7. [#49246](#49246) adds a performance measurement suite. It identifies where search time is spent and provides a reliable way to evaluate performance improvements. 8. [#49247](#49247) delivers the main performance improvement. App results are scored in parallel and expensive work no longer blocks rendering, while the final result order remains unchanged. 9. [#49249](#49249) prevents misleading results from flashing when a search begins. For one or two character searches, weak fuzzy app matches remain hidden until the query is specific enough to produce useful results. > [!WARNING] > These PRs should be merged in LIFO order starting with #49249 with this PR being the last. --------- Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

For 1-2 char queries a huge fraction of the app catalog fuzzy-matches into the single Fuzzy tier, where within-tier order is decided almost entirely by frecency, so the most-frecent app can float to rank 1 on a weak mid-word subsequence match - the "wrong when I start typing" flash. Gate the render path only: while the query is ultra-short, show apps at or above the word-boundary tier and withhold the low-confidence fuzzy tail until the query can discriminate. This slices the already-scored, already-sorted app array and never re-scores or reorders, so the settled order of a 3+ char query is byte-identical and the Phase 7b equivalence guarantee is untouched. Applying the same gate to a late async app batch also stabilizes short-query fold-in. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

…metry count Fix #2: publish the app query length atomically with _filteredApps under lock(commands) and gate the early-frame render path on that published field instead of the live SearchText. SearchText is advanced on the UI thread ahead of publish, so on a 2-to-3 char transition a throttled refresh could render the stale 2-char app array while the live length read 3, no-op the gate, and briefly expose the fuzzy tail 7c withholds. _filteredAppsQueryLength is reset in ClearResults alongside _filteredApps. Fix #3: count the gated visible apps for the settled-search telemetry when the published length is in the ultra-short gate window, so NoResults / result count reflect what is actually shown rather than the ungated app array. Added GatedVisibleAppCount, kept the 7b post-lock telemetry structure. Tests: added coverage that the gate decides on the supplied published length (not live text) and that the telemetry count equals the gated visible count. EarlyFrameRelevanceTests and 7b ScoringParallelEquivalenceTests both stay green. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

The off-lock scoring phase pins every input into immutable locals so a concurrent keystroke cannot mix two generations into one pass. PrecomputeQuery still read the live SearchText field, which a newer keystroke may already have advanced, scoring the pass against a different query than it was launched for. Precompute from the snapshotted newSearch parameter instead. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com> Copilot-Session: 5459b847-afeb-4163-a803-977759dd92df

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com> Copilot-Session: 92905c83-2de6-449c-b4a1-a08003fe2576

Mike Griese (zadjii-msft)

yea code's fine. Just nitpicking over phrasing again. Still feels like the agents are saying more that the right number of words

Mike Griese (zadjii-msft) · 2026-08-06T21:10:15Z

+
+    // The array is sorted descending by packed score and the tier lives in that score's high bits,
+    // so everything at or above a tier forms a contiguous prefix that one forward scan can find.
+    internal static int HighConfidenceAppPrefixLength(IReadOnlyList<RoScored<IListItem>> scored, RankTier minTier)


<nit>
HighConfidenceAppPrefixLength makes me think that this is string handling. Maybe from the word Prefix. Maybe GetHighConfidenceAppsCount or something like that?

Mike Griese (zadjii-msft) · 2026-08-06T21:11:11Z

+    // Keeps only the letter-relevant apps while the query is short, and returns a longer query's
+    // results untouched. It slices a contiguous prefix of the already-scored, already-sorted array,
+    // so it can never re-score or reorder anything.
+    internal static IList<RoScored<IListItem>>? ApplyShortQueryAppGate(


FilterAppsForShortQueries might be better? The word gate really feels like a lock or synchronization primitive, not a filter.

Mike Griese (zadjii-msft) · 2026-08-06T21:14:45Z

+        // Hold back the weak fuzzy tail while the query is still short, so a frecency-floated
+        // match can't surface at the top mid-typing.


Suggested change

// Hold back the weak fuzzy tail while the query is still short, so a frecency-floated

// match can't surface at the top mid-typing.

// Remove fuzzy-only app matches for short queries. This prevents

// frecency-boosted (but otherwise weakly matching) apps from appearing

// at the top while typing

"hold back" seemed weird? I'm not sure mine is better

Mike Griese (zadjii-msft) · 2026-08-06T21:18:14Z


    private int AppResultLimit => AllAppsCommandProvider.TopLevelResultLimit;

+    // ===== Early-frame relevance =====


Suggested change

// ===== Early-frame relevance =====

I don't think that adds any information? "Early-frame" doesn't feel like a real term we've well defined.

Mike Griese (zadjii-msft) · 2026-08-06T21:29:46Z

    private int AppResultLimit => AllAppsCommandProvider.TopLevelResultLimit;

+    // ===== Early-frame relevance =====
+    // Longest query we still hold back the weak fuzzy app tail for, since one letter matches most


<nit> something about the phrase "fuzzy tail" just doesn't immediately click for me.

Can we rephrase this comment to something like

// Longest query to filter fuzzy app matches on. This prevents weak (fuzzy) app matches on short queries.

Mike Griese (zadjii-msft) · 2026-08-06T21:32:02Z

+{
+    public TestContext TestContext { get; set; } = null!;
+
+    private sealed partial class CatalogItem : IListItem, IPrecomputedListItem


github-actions Bot added the Product-Command Palette Refers to the Command Palette utility label Jul 10, 2026

Michael Jolley (michaeljolley) mentioned this pull request Jul 11, 2026

[CmdPal] Replace main-page magic-number scoring with a principled tiered ranker #49189

Merged

Michael Jolley (michaeljolley) requested review from Jiří Polášek (jiripolasek) and Mike Griese (zadjii-msft) July 11, 2026 21:42

Michael Jolley (michaeljolley) marked this pull request as ready for review July 11, 2026 21:42

Michael Jolley (michaeljolley) mentioned this pull request Jul 11, 2026

[CmdPal] Phase 7b: parallel + off-lock search scoring throughput #49247

Open

Michael Jolley (michaeljolley) self-assigned this Jul 11, 2026

Gordon Lam (yeelam-gordon) mentioned this pull request Jul 12, 2026

[review-mirror] microsoft/PowerToys#49249 yeelam-gordon/PowerToys#21

Open

Michael Jolley (michaeljolley) force-pushed the dev/mjolley/search-perf-early-frame branch from ab0fa5d to ff39910 Compare August 3, 2026 17:57

Michael Jolley (michaeljolley) force-pushed the dev/mjolley/search-perf-early-frame branch from ff39910 to b575b5a Compare August 3, 2026 18:39

Michael Jolley (michaeljolley) force-pushed the dev/mjolley/search-perf-early-frame branch from b575b5a to 20935d9 Compare August 3, 2026 22:22

Michael Jolley (michaeljolley) force-pushed the dev/mjolley/search-perf-early-frame branch 2 times, most recently from dcd842c to 6a98532 Compare August 4, 2026 17:14

Michael Jolley (michaeljolley) force-pushed the dev/mjolley/search-perf-early-frame branch from 6a98532 to 0d5413b Compare August 4, 2026 20:26

Michael Jolley (michaeljolley) force-pushed the dev/mjolley/search-perf-early-frame branch from 0d5413b to 735d1c1 Compare August 4, 2026 21:45

Michael Jolley (michaeljolley) force-pushed the dev/mjolley/search-perf-early-frame branch from 735d1c1 to c3b6234 Compare August 6, 2026 18:14

Michael Jolley (michaeljolley) and others added 2 commits August 6, 2026 14:57

Shorten phase 7c comments

e1d4522

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com> Copilot-Session: 92905c83-2de6-449c-b4a1-a08003fe2576

Michael Jolley (michaeljolley) force-pushed the dev/mjolley/search-perf-early-frame branch from c3b6234 to e1d4522 Compare August 6, 2026 20:08

Mike Griese (zadjii-msft) reviewed Aug 6, 2026

View reviewed changes

Navigation Menu

Uh oh!

[CmdPal] Phase 7c: early-frame relevance (withhold weak fuzzy tail on ultra-short queries) - #49249

[CmdPal] Phase 7c: early-frame relevance (withhold weak fuzzy tail on ultra-short queries)#49249
Michael Jolley (michaeljolley) wants to merge 4 commits into
dev/mjolley/search-perf-throughputfrom
dev/mjolley/search-perf-early-frame

Michael Jolley (michaeljolley) commented Jul 10, 2026 •

edited

Loading

Uh oh!

Michael Jolley (michaeljolley) commented Aug 3, 2026

Uh oh!

azure-pipelines Bot commented Aug 3, 2026

Uh oh!

Mike Griese (zadjii-msft) left a comment

Uh oh!

Mike Griese (zadjii-msft) Aug 6, 2026

Uh oh!

Mike Griese (zadjii-msft) Aug 6, 2026

Uh oh!

Mike Griese (zadjii-msft) Aug 6, 2026

Uh oh!

Mike Griese (zadjii-msft) Aug 6, 2026

Uh oh!

Mike Griese (zadjii-msft) Aug 6, 2026

Uh oh!

Mike Griese (zadjii-msft) Aug 6, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

		// Hold back the weak fuzzy tail while the query is still short, so a frecency-floated
		// match can't surface at the top mid-typing.

-        // Hold back the weak fuzzy tail while the query is still short, so a frecency-floated
-        // match can't surface at the top mid-typing.
+        // Remove fuzzy-only app matches for short queries. This prevents
+        // frecency-boosted (but otherwise weakly matching) apps from appearing
+        // at the top while typing


		private int AppResultLimit => AllAppsCommandProvider.TopLevelResultLimit;

		// ===== Early-frame relevance =====

Uh oh!

Conversation

Michael Jolley (michaeljolley) commented Jul 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What's going on

The plan

Notes for reviewers

Uh oh!

Michael Jolley (michaeljolley) commented Aug 3, 2026

Uh oh!

azure-pipelines Bot commented Aug 3, 2026

Uh oh!

Mike Griese (zadjii-msft) left a comment

Choose a reason for hiding this comment

Uh oh!

Mike Griese (zadjii-msft) Aug 6, 2026

Choose a reason for hiding this comment

Uh oh!

Mike Griese (zadjii-msft) Aug 6, 2026

Choose a reason for hiding this comment

Uh oh!

Mike Griese (zadjii-msft) Aug 6, 2026

Choose a reason for hiding this comment

Uh oh!

Mike Griese (zadjii-msft) Aug 6, 2026

Choose a reason for hiding this comment

Uh oh!

Mike Griese (zadjii-msft) Aug 6, 2026

Choose a reason for hiding this comment

Uh oh!

Mike Griese (zadjii-msft) Aug 6, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Michael Jolley (michaeljolley) commented Jul 10, 2026 •

edited

Loading