PickerSearch.md

status	in-force
summary	The install picker's search ranks with a FuzzyMatch port over the card fields, the CMS tags, the operator types a tool stands in for and a docs word index each manifest row carries as `search`; a query shows one ranked list, and a Compact toggle draws title-only tiles.
since	v3.2.63 (2026-10-05)
verified	2026-10-03
skill	fns-packaging

Install picker search and compact view

Owner feedback on the install picker (/get/, the installer's served page, the console's Install tab), 2026-10-03: "there are just so many tools and features". The search was a plain substring over what a card shows (name, description, category, surfaces), so someone typing "midi" or "undo" never found a tool whose card did not happen to say it, and every hit sat in its category section, so a search scattered its answers down the page.

Two changes, both in packaging/configurator/index.html:

• Search goes deeper and ranks. It matches the card fields, the CMS tags, the operator types a tool is an alternative for, and the words of the tool's documentation. While a query is present the list is ONE ranked grid, best match first.
• Compact view. A toggle in the filter bar draws title-only tiles.

What a package answers to

FNSSearch.prepare(p, labelOf) turns a manifest row into fields; each field has a weight that is ADDED to the FuzzyMatch tier of a hit in it, so a title hit outranks the same hit in a description:

Field	Weight	Fuzzy?
title, name (minus FNS_), family op_label	0	yes
each family search_words tag (the CMS tags), each alternatives_for type	0.3	yes
category, family op_group, each surface's label, author name	0.4	strict
description, family summary, trial + pricing summary, each variant id + summary	0.5	strict
docs head words (search.head)	0.6	word start only
docs body words (search.body)	0.8	word start only

Ranking is scripts/shared/FuzzyMatch.py's, ported line for line into the page as FNSFuzzy (EXACT, PREFIX, WORD, SUBSTRING, then the fuzzy INITIALS, TYPO, SUBSEQUENCE). A token takes its best field; a row's tier is its worst token's weighted tier; ties go to the tighter hit, then the title.

The rules that keep it honest:

• Every weight is under 1, so a strict hit anywhere (a docs word included, at most 2.8; a strict substring, at most 3.5) still ranks above any fuzzy one (4 and up). This is the FuzzyMatch promise that the fuzzy tiers only ever fill BELOW the literal hits.
• Only short name fields take fuzzy hits. A typo or a subsequence over a description or a docs page matches almost anything.
• A token under three letters matches word starts only: no substring, no docs. "ui" inside "quick" put every Quick* tool on the list, and "ab" matched 26 tools through table, label and absolute.
• A subsequence needs four letters. "lag" was a subsequence of SimpleSceneChanger.
• Hotkeys are not searched. Nearly every binding carries alt or ctrl, so "alt" listed 41 tools with CustomParTools above AltSelect.

The docs index: search on each manifest row

packaging/search_index.py (pure Python, no TD imports) reads packaging/docs/<Package>.md and returns {"head": "...", "body": "..."}: two strings of space-joined, sorted, deduplicated words.

• head: the frontmatter summary and feature names, the ## headings, the **bold** terms. What the doc says the tool is about.
• body: every other prose word, minus the head.
• Folded like a query. Words come from FuzzyMatch.words, so the page compares like with like: ChopBank gives chop, bank and the compound chopbank; zero-below gives zero and zerobelow; colour folds to color. An acronym plural is one word (CHOPs -> chops, never cho + ps), trailing digits go (Fader1 -> fader).
• Dropped: words under three letters, stopwords, fenced and inline code, link targets, HTML tags, URLs, every word the card already answers to (title, name, description, family tags and summary; those fields score higher anyway), and any word in more than 30% of all docs ("parameter", "network", "touchdesigner": they say nothing about one tool).
• Matched by word start only. A token hits a docs word it starts, never one it merely sits inside, and never fuzzily. The card shows in docs: undo when a docs word carried a token no card field did, so a deep hit explains itself.

build_manifest.Build() loads the module by path (the build is exec'd into TD's namespace, so there is no package to import from), builds the index over every row once, and writes search presence-style: a row whose doc adds nothing carries no key. A preview row carries none either: the published manifest names a preview and describes it in one line (docs/PreviewPackages.md, Limits), and only the creator's picker shows it. Build() reports the cost as search_bytes.

Size budget

Measured 2026-10-03 over 98 docs: 80 KB raw, 26 KB gzipped, against a 153 KB manifest. tests/test_search_index.py fails above 100 KB raw or 35 KB gzipped. The cheapest further cut, if it is ever needed, is the common-word share (20% saves about 5 KB); CamSequencer's doc alone is a tenth of the index.

Where the field travels

• Build() writes it into manifest.json, configurator/manifest.js and the standalone, so it rides the store manifest the updater downloads and the installer serves to the in-TD picker (/manifest.js) without a route of its own.
• The updater reads package rows with .get() and indexes them by name; nothing validates manifest keys strictly (checked 2026-10-03), so an unknown key is ignored.
• /get/: website/tools/build-site.mjs back-fills search from the repo manifest into the baked page (the docs are newer than any published release, the same reason the catalog's description wins), and the page's runtime merge keeps the baked search over a live manifest that lacks it.
• A manifest without search still searches every other field.

The results

• A query with at least one token renders one ranked grid (.grid.ranked). Each card names its category, since there are no section headings. The sections and their fold state come back when the field is cleared.
• Enter in the search ticks the top hit, through the card's own checkbox so a Patreon tool still asks; it never unticks.
• The category rail, under a search, scrolls to that category's best hit instead of a section that is not drawn.
• The results line, the filter pills, the access counts and Select/Clear shown act on exactly what is listed (shownTools), as before.

Compact view

A Compact toggle in .bar-filters, off by default, remembered per viewer in localStorage (fns.compact, wrapped in try/catch like every key on the page). It is a view, not a filter: every card stays, only the words go. It toggles the dense class on #list; nothing re-renders.

• Kept: the checkbox, the title, New, installed / will install / will be removed, the Patreon chip, a trial chip, a placed tool's undo, and under a search the category and the in docs: line.
• Hidden: the description, the byline, the docs and site links, the hint chips, the where-it-lands chips, the variant chips, and Place on a card that is not placed.
• The description is the tile's hover: the card's data-tip already carries it.

The picker-only CSS sits outside the FNS:UIBASE markers, which sync_base.py regenerates from base.css.

Rejected

• Fetching the docs at runtime. The page served from inside TouchDesigner has no site to fetch from, and the standalone works offline.
• A separate index file. One more fetch per flavor, and the bucket sends no CORS headers, so a 127.0.0.1 page would need its own installer relay. On the manifest row it rides every existing path for free.
• Substring over the docs text. Folding joins the words, so a substring matches across word boundaries ("lag" in "dial agent"), and a long text holds nearly every short string.
• Fuzzy over the docs. A subsequence of a few hundred words matches anything.

Rolling it out

The page change works the day it lands: the ranked list and Compact need no new data. The docs words reach a picker with the next release's Build(). The repo manifest was not regenerated when this landed: a mid-cycle Build() is release-shaped. On 2026-10-03 it added the two unreleased packages (PaneSearchRegistry, SearchFix), whose whatsnew the version-carry rule would then freeze from that day's notes. It also rewrote twenty rows with live drift (portability, hotkeys, cooking) and re-rendered the surface icons. The served picker inside TouchDesigner is a snapshot build_installer embeds into FNS_Installer/configurator_html; BuildBootstrap() refreshes it at release, and EnsureDevRails() refreshes the dev root's copy on demand.

Tests

• tests/test_search_index.py: word folding and drops, head and body, the real index against the committed manifest, the size budget, and that Build() and the site carry the field.
• tests/test_picker_search.py: evaluates the page's own FNS:SEARCH block in node from both shells. FNSFuzzy against FuzzyMatch.py over the palette corpus (every tier), the ranking table on the real catalog (math chop, midi fader, a CMS tag, typos, a docs-only word, short queries), the strict-above-fuzzy invariant, and the page wiring.
• tests/test_foreign_packages.py and tests/test_tier_variants.py pin that a trial and a variant summary stay searchable.