How to Add Full-Text Search to a Site with Elasticsearch

How it works in short

Elasticsearch is a separate search index next to your main database. The database stays the source of truth; Elasticsearch gets a copy of the searchable fields, prepared by analyzers that split text into tokens, lowercase them and reduce words to their base form. Search quality depends mostly on three things: the mapping, the sync process and the query.

Step 1. Design the index and analyzers

Store each language in its own field, so each gets the right analyzer.

• Russian: the built-in russian analyzer handles stop words and stemming ("телефоны" matches "телефон").
• Uzbek: there is no built-in Uzbek analyzer. A practical baseline is a custom analyzer that lowercases text and removes the different apostrophe characters (‘ ’ ʻ ʼ ' and backtick), because users type o‘zbek, o'zbek and ozbek interchangeably.

PUT /products
{
  "settings": {
    "analysis": {
      "char_filter": {
        "uz_apostrophes": {
          "type": "pattern_replace",
          "pattern": "[‘’ʻʼ'`]",
          "replacement": ""
        },
        "ru_yo": { "type": "mapping", "mappings": ["ё => е", "Ё => Е"] }
      },
      "analyzer": {
        "uz_text": {
          "type": "custom",
          "char_filter": ["uz_apostrophes"],
          "tokenizer": "standard",
          "filter": ["lowercase"]
        },
        "ru_text": {
          "type": "custom",
          "char_filter": ["ru_yo"],
          "tokenizer": "standard",
          "filter": ["lowercase", "russian_stop", "russian_stemmer"]
        }
      },
      "filter": {
        "russian_stop": { "type": "stop", "stopwords": "_russian_" },
        "russian_stemmer": { "type": "stemmer", "language": "russian" }
      }
    }
  },
  "mappings": {
    "properties": {
      "title_ru": {
        "type": "text", "analyzer": "ru_text",
        "fields": { "suggest": { "type": "search_as_you_type", "analyzer": "ru_text" } }
      },
      "title_uz": {
        "type": "text", "analyzer": "uz_text",
        "fields": { "suggest": { "type": "search_as_you_type", "analyzer": "uz_text" } }
      },
      "description_ru": { "type": "text", "analyzer": "ru_text" },
      "description_uz": { "type": "text", "analyzer": "uz_text" },
      "category": { "type": "keyword" },
      "price": { "type": "integer" },
      "updated_at": { "type": "date" }
    }
  }
}

Uzbek is agglutinative: without a stemmer, "kitoblar" and "kitob" are different tokens. Typo tolerance and prefix search (below) partly cover this. Test with real queries from your users, and check token output with the _analyze API.

Use keyword for fields you filter or aggregate on (category, brand, status) and text for fields you search in.

Step 2. Sync data from the main database

Practical rules:

• Use the Bulk API for batches, not one request per document.
• Handle deletes explicitly, for example with a deleted_at column.
• Read and write through an index alias. For a full rebuild, create a new index, fill it, then switch the alias with no downtime.

Step 3. Typo tolerance

fuzziness: "AUTO" allows a small number of character edits depending on word length.

GET /products/_search
{
  "query": {
    "multi_match": {
      "query": "telefn samsung",
      "fields": ["title_ru^3", "title_uz^3", "description_ru", "description_uz"],
      "fuzziness": "AUTO",
      "prefix_length": 1
    }
  }
}

prefix_length: 1 requires the first letter to match, which reduces noise and speeds up the query. The ^3 boost makes title matches rank above description matches.

Step 4. Autocomplete

The search_as_you_type field creates helper subfields for prefix matching. Query them with bool_prefix:

GET /products/_search
{
  "size": 5,
  "_source": ["title_ru", "title_uz"],
  "query": {
    "multi_match": {
      "query": "smartf",
      "type": "bool_prefix",
      "fields": [
        "title_ru.suggest", "title_ru.suggest._2gram", "title_ru.suggest._3gram",
        "title_uz.suggest", "title_uz.suggest._2gram", "title_uz.suggest._3gram"
      ]
    }
  }
}

On the frontend, send the request after a short debounce and cancel outdated requests.

Common mistakes

• One text field with the default analyzer for all languages.
• Exposing Elasticsearch directly to the browser. Keep it behind your backend and never open it to the internet without authentication.
• Changing a mapping in place. Most mapping changes require a reindex into a new index.