Search solutions for static sites

Options for adding search to documentation, blogs and static-site gardens —
client-side (index shipped to the browser) vs server-side (self-hosted
engine). Researched 2026-09-13 for garden.h; resource numbers are indicative,
verify before committing. Elasticsearch excluded by policy (JVM, GB-class
RAM for what we need).

Requirement scope: English + Polish as targets. Czech/Slovak are tracked
only as reference implementations — how full West-Slavic language support
looks when adding Polish elsewhere. Shortlist under that scope:
Pagefind (EN porter2 + PL Snowball, zero backend) and ZBSearch
(Orama fork by the original team; EN built-in; PL via a custom stemmer —
see below) client-side; Meilisearch (explicit eng+pol locales,
typo tolerance) or Typesense (EN+PL stemming, ~30 MB idle) server-side.
lunr-languages works but downloads the whole index.

Client-side (no backend)

• FlexSearch — what Quartz search uses. ~5 KB, index built at
  runtime from a JSON corpus. Fast prefix matching; no stemming, no
  diacritics folding. Fine ≤ a few hundred notes.
• lunr + lunr-languages — classic SSG search (Hugo/MkDocs heritage).
  Full index downloaded upfront (bandwidth grows linearly — bad for large
  sites). Polish: ✅ (lunr-languages ships a PL stemmer + stopwords);
  Czech/Slovak: ❌.
• Pagefind — Rust CLI indexes built HTML, ships chunked WASM index;
  searches load only the needed chunks (~100 KB payload for most sites,
  MDN-scale < 300 KB). Zero config, per-section results, filters.
  Multilingual via lang attribute; stemming follows Snowball →
  Polish: ✅ (added in the 1.x Snowball upgrade), Czech: ✅, Slovak: ❌.
  Best “just works at scale” option; no quartz-community plugin (see
  Quartz search).
• Orama / ZBSearch (~2 KB; Orama formerly Lyra) — full-text + vector +
  hybrid in browser or server, BM25, typo tolerance. Prefers ZBSearch:
  fork by the original Orama team (ex co-founder/CTO Michele Riva + the
  whole engineering team) after the VC fallout — same engine, no business
  incentives, actively developed (v4.0.0). 32 languages with tokenizer +
  stopwords + stemmer: Czech ✅ (port of Lucene’s CzechStemmer, Dolamic &
  Savoy 2009 — suffix tables + palatalization, dependency-free), Slovak ✅
  (ZBSearch-only: hand-written light stemmer; explains why the Czech one is
  not a stand-in and how diacritic folding interacts with stemming — the
  blueprint for a Polish stemmer), Slovenian ✅, Polish ❌ in both.
  Gotcha: Orama’s sk code is Sanskrit, not Slovak (ZBSearch fixed
  this with svk). Custom stemmers are first-class:
  Stemmer = (word) => string passed to components.tokenizer; most
  built-ins are Snowball-compiled JS, so a Polish stemmer is
  snowball compiler → polish.sbl → JS plug-in. Strong middle-ground
  option for EN+PL.

Server-side (self-hosted engines)

• Meilisearch — Rust, single binary, typo-tolerant, ~instant. Charabia
  tokenizer; locales setting (ISO-639-3) explicitly lists pol, ces AND
  slk — the only engine surveyed with explicit Slovak support. Hybrid
  keyword+vector search built in. Homelab-fit: container or binary,
  moderate RAM (~100–200 MB class).
• Typesense — C++, single binary, ~30 MB idle, RAM-mapped, very fast.
  locale accepts any ISO 639-1 incl. pl/cs/sk → ICU normalization +
  collation; Snowball stemming → Czech ✅, Polish ✅ (Snowball PL exists),
  Slovak ❌ (no Snowball algorithm — gets normalization only). Dictionary
  stemming (stem_dictionary) as escape hatch.
• Algolia — SaaS only (no self-host). Free 10k searches/mo, 50k records;
  OSS projects get 200k/mo with attribution. Excellent DX (InstantSearch
  UI), but data leaves your infrastructure — conflicts with the self-hosting
  premise. Reference point, not a candidate.
• QMD — for the agent/wiki side (not website visitors): BM25 +
  vectors + LLM rerank in one local binary. Different problem, same family.

Language support matrix (PL / CS / SK)

Engine	Polish	Czech	Slovak
Snowball (reference)	✅	✅	❌
lunr-languages	✅	❌	❌
Pagefind (Snowball)	✅	✅	❌
Orama	❌	✅	❌
Meilisearch locales	✅	✅	✅
Typesense (ICU+Snowball)	✅	✅	~ (norm. only)
FlexSearch (any lang)	❌ no stemming at all	❌	❌

Czech is the best-represented West-Slavic language (official Snowball
algorithm + Orama full support) — for our EN+PL scope it serves purely as a
reference: the reference implementation pattern to copy when adding Polish
support anywhere. Slovak is covered only by Meilisearch.

Choosing

• Small garden/blog (≤ ~500 pages): client-side. Pagefind if stems/multilingual
  matter; stock FlexSearch (Quartz search) if prefix search suffices.
• Docs portal / work: Meilisearch (PL+CS+SK explicit, typo tolerance,
  hybrid) or Typesense (leanest RAM) behind the reverse proxy.
• Inflected-Polish matching on a static site without a server: pre-stem the
  corpus at build time — see Polish stemming and lemmatization.

Cross-refs: Quartz search, Polish stemming and lemmatization,
QMD, RAG, Local LM roadmap.