For AI agents: this page is also available as Markdown at /docs/hebrew.md. Every docs page is: add .md to its URL. llms.txt lists them all.

Hebrew

Hebrew attaches prepositions, conjunctions and the definite article to the word that follows (לבית, "to the house"), inflects nouns and verbs heavily, and is usually written without vowel points, with extra ו and י letters standing in for them. A plain word index treats each of these forms as a different word. The hebrew analyzer looks every word up in a Hebrew dictionary of 337,000 words and indexes its lemmas, so a search for בית finds לבית, ובבית and הבית, and a search for ספר finds הספרים.

{"fields": {"body_he": {"type": "text", "analyzer": "hebrew"}}}

It is a port of HebMorph, the analyzer of the Elasticsearch Hebrew plugin, with its dictionary built in. No license key or separate dictionary download is needed.

#What gets indexed

For each word the analyzer removes niqqud (vowel points), then indexes two kinds of terms at the same position:

• The exact form, with a $ suffix: הספרים$. It matches only that spelling.
• Every lemma the dictionary gives for the word, with its prefixes removed: ספר. A word with several readings gets several lemmas: שמלות is "dresses" (שמלה) or "that from" (ש + מלות, lemma מל).

For the sentence הילדים הלכו לבית הספר ("the children went to school"):

So ילד, ילדים and הילד all find this document, and so do בית and ספר.

Other words:

• Spelling without niqqud is tolerated: אמא is looked up as אימא too, and an extra or missing ו or י still finds the lemma. Words over 20 letters are not tolerated.
• Words not in the dictionary, such as most names, are their own lemma: book gives book$ and book.
• Acronyms written with gershayim (צה"ל) and words with a geresh (ג'ירפה) stay one word.
• Words starting with a digit are indexed as they are, with no exact form and no lemmas.
• Case and Unicode width are folded first, so Latin words match in any case.

#Exact matches

A query with the same analyzer produces the exact form and the lemmas of each query word, so it matches every inflection. Documents that use the query's exact spelling also match its $ term, which usually ranks them higher. To match only the exact spelling, search the $ form with a term query, which is not analyzed:

{"language": "es", "query": {"term": {"body_he": "הספרים$"}}}

Write the term without niqqud, as it is indexed.

#Prefix, wildcard and fuzzy terms

hebrew_lemmatizer does not run on prefix, wildcard and fuzzy terms (lemmatizing part of a word means nothing), so they match the indexed terms as written: ספר* matches the lemma ספר and exact forms that start with ספר, but not הספרים$. The terms are still case folded.

#The filter and the presets

To combine Hebrew with other filters, write out the chain, keeping hebrew_lemmatizer last:

{
  "type": "text",
  "analyzer": {
    "char_filters": ["html_strip"],
    "tokenizer": "unicode_word",
    "filters": ["nfkc_cf", "fold", {"name": "hebrew_lemmatizer", "params": {"suffix": "$"}}]
  }
}

#Migrating from Elasticsearch

Mappings that name the Elasticsearch Hebrew plugin's hebrew or hebrew_query analyzer get the WarpSearch analyzers of the same names. hebrew_exact maps to hebrew and hebrew_query_light to hebrew_query, each with a warning: search the $ forms for exact matches, as above. Results match the Java analyzer run over the same dictionary, which WarpSearch tests against. See Through the compatible APIs.

View this page as Markdown