From a9d9634678bb3c81a11a0080bc244ace95fab036 Mon Sep 17 00:00:00 2001
From: teknium1 <127238744+teknium1@users.noreply.github.com>
Date: Sat, 8 Aug 2026 05:39:32 -0700
Subject: [PATCH] perf(docs-site): split search index by section and stop
 indexing code blocks
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The client-side search index had grown to 16.3 MB raw / ~4.5 MB wire
(4,500+ docs), giving 25-30 s of blank UI before first results — users
read the dead window as 'search is broken entirely' (the engine itself
returned correct results).

Three levers, all config-only:

- searchContextByPaths: per-section index chunks. Searching from any
  docs page now fetches only that section's index (e.g. Reference:
  1.4 MB / 387 KB gz) instead of the 16 MB monolith. The search page
  gains a section dropdown; landing-page searches still cover
  everything via useAllContextsWithNoSearchContext.
- ignoreCssSelectors: ['pre']: fenced code blocks no longer indexed.
  YAML/shell samples were generating huge high-cardinality lunr token
  dictionaries; inline code in prose stays searchable.
- ignoreFiles: user-stories excluded (527 index docs of scraped
  community quotes rendered by a React component).

Measured (npm run build, both locales):
- root 'Everywhere' index: 16.3 MB -> 13.2 MB raw (4.42 -> 3.57 MB gz)
- per-section indexes: 0.4-8.3 MB raw (103 KB-2.2 MB gz)
- zh-Hans root: 14.6 -> 12.6 MB raw
- local serve: first dropdown results in ~175-205 ms for both scoped
  and Everywhere queries ('telegram' -> 8 options, search page -> 100)
---
 website/docusaurus.config.ts | 24 ++++++++++++++++++++++++
 1 file changed, 24 insertions(+)

diff --git a/website/docusaurus.config.ts b/website/docusaurus.config.ts
index a10987be10269..0a5792866d59b 100644
--- a/website/docusaurus.config.ts
+++ b/website/docusaurus.config.ts
@@ -61,7 +61,31 @@ const config: Config = {
         ignoreFiles: [
           /^user-guide\/skills\/bundled\//,
           /^user-guide\/skills\/optional\//,
+          // Community-quote collage page: 527 index documents of scraped
+          // testimonials that pollute results and bloat the index. The page
+          // renders a React component; its text isn't reference material.
+          /^user-stories/,
         ],
+        // Split the lunr index into per-section chunks so the browser only
+        // fetches+hydrates the index for the section being searched instead
+        // of one monolithic multi-MB file (16.3 MB by Aug 2026 — 25-30 s of
+        // blank UI on first search). Searches started outside these paths
+        // (e.g. the docs landing page) fall back to all contexts combined.
+        searchContextByPaths: [
+          { label: 'User Guide', path: 'user-guide' },
+          { label: 'Developer Guide', path: 'developer-guide' },
+          { label: 'Guides', path: 'guides' },
+          { label: 'Reference', path: 'reference' },
+          { label: 'Getting Started', path: 'getting-started' },
+          { label: 'Integrations', path: 'integrations' },
+        ],
+        useAllContextsWithNoSearchContext: true,
+        // Don't index fenced code blocks (<pre>). Config/YAML/shell examples
+        // generate huge high-cardinality token dictionaries in lunr (the
+        // biggest single contributor to index size) and nobody searches for
+        // a literal line of a code sample. Inline <code> (command and config
+        // key names in prose) stays indexed.
+        ignoreCssSelectors: ['pre'],
         // Exact-or-prefix matching only (default is edit distance 1).
         // With fuzzy distance 1, "keet" matched "meetings"/"keep" (one
         // edit away after stemming), and multi-word typo queries against