Multilingual analysis

One index, twenty-plus languages, and byte-identical analysis everywhere — Node and every browser.

Turn on the multilingual profile in the build config:

{ "analysis": { "profile": "multi-v1" } }

or pin languages explicitly:

{ "analysis": { "profile": "multi-v1", "languages": ["en", "fr", "de"], "languageField": "lang" } }

What the profile does

• Per-document language — detected from content, or read from languageField when your data already knows.
• Light stemmers and stopwords for 20+ languages, tuned for recall without aggressive over-stemming.
• Script-aware folding — Latin, Greek, Cyrillic, Arabic, Hebrew, and Devanagari diacritics and forms normalize predictably, so Émile matches emile.
• CJK bigram tokenization — dictionary-free and deterministic, so Chinese, Japanese, and Korean text is searchable without shipping segmentation dictionaries.
• Cross-language query expansion — a query analyzed under several plausible languages picks the plan with the best index evidence, so a French query against a mixed corpus finds French stems without an explicit language switch.

The determinism rule

The entire analysis profile is frozen into the manifest at build time, and the runtime reconstructs the exact same analyzer from it. Nothing depends on ICU or platform locale tables — the same text tokenizes to the same terms in Node 22, Chrome, Safari, and Firefox. This is what makes a static index safe: query analysis can never drift from index analysis, even across browser versions and years.

The practical consequence: incremental updates and shards must be built with the same analysis config as their base — the builder enforces this and refuses mismatched profiles rather than producing silently incomparable terms.