Fuzzy Search Implementation for Web Applications

Fuzzy Search Implementation for Web Applications

A user types "наушниик" — empty result. 70% of those visitors leave for competitors. Fuzzy search fixes typos and returns relevant products. Over the years, we've implemented fuzzy search for more than 30 e-commerce and catalog projects. We choose the engine that fits your stack and load: pg_trgm, Meilisearch, or Elasticsearch. With proper tuning, conversion grows by 15–25%, and maintenance costs drop — our cases confirm this.

For example, for an online home appliance store, we reduced the rate of empty results from 25% to 3% by switching to Meilisearch. Conversion increased by 22%. Response time dropped from 200 ms to 4 ms. Order a pilot project — we'll test on your data.

What distance algorithms are used?

Levenshtein distance — the minimum number of insertions, deletions, and substitutions to change one string into another, described on Wikipedia. Damerau-Levenshtein distance adds transposition (swapping adjacent characters). For Russian, it's preferable: "наушники" → "наушинки" is one transposition instead of two operations. In practice, we use Damerau-Levenshtein. The choice affects quality: Damerau-Levenshtein yields 10% fewer misses for Russian queries.

PostgreSQL: pg_trgm

The pg_trgm extension works with trigrams and requires no external services. It's the simplest solution if your stack already includes PostgreSQL.

CREATE EXTENSION IF NOT EXISTS pg_trgm;
CREATE INDEX idx_products_title_trgm ON products USING GIN (title gin_trgm_ops);
CREATE INDEX idx_products_description_trgm ON products USING GIN (description gin_trgm_ops);
SET pg_trgm.similarity_threshold = 0.3;
SELECT id, title, similarity(title, 'наушниик') AS sim
FROM products
WHERE title % 'наушниик'
ORDER BY sim DESC
LIMIT 10;
-- Combine fuzzy with full-text search
SELECT p.id, p.title, p.price,
       greatest(similarity(p.title, 'беспродные наушники'), ts_rank(p.search_vector, plainto_tsquery('russian', 'беспродные наушники'))) AS relevance
FROM products p
WHERE p.title % 'беспродные наушники' OR p.search_vector @@ plainto_tsquery('russian', 'беспродные')
ORDER BY relevance DESC
LIMIT 20;

The % operator uses the GIN index. The similarity_threshold of 0.3 is liberal; 0.5 is strict. For short queries, choose the lower bound. In practice, we recommend starting at 0.3 and adjusting through A/B tests: raising the threshold to 0.5 reduces false positives but may miss some relevant results. Infrastructure savings with pg_trgm amount to up to 40% compared to external engines.

Meilisearch — Dedicated Fuzzy Engine

Meilisearch is written in Rust and supports typo tolerance out of the box. It's specifically designed for fast fuzzy search and requires little configuration.

import meilisearch
client = meilisearch.Client('http://localhost:7700', 'your-master-key')
index = client.index('products')

# Index settings
index.update_settings({
    'searchableAttributes': ['title', 'brand', 'description', 'tags'],
    'filterableAttributes': ['category_id', 'status', 'price', 'brand'],
    'sortableAttributes': ['price', 'created_at', 'popularity'],
    'rankingRules': ['words', 'typo', 'proximity', 'attribute', 'sort', 'exactness'],
    'typoTolerance': {
        'enabled': True,
        'minWordSizeForTypos': {
            'oneTypo': 5,
            'twoTypos': 9
        },
        'disableOnWords': ['iPhone', 'iPad'],
        'disableOnAttributes': ['sku', 'barcode'],
    },
    'pagination': {
        'maxTotalHits': 10000
    },
})

# Batch indexing
batch_size = 1000
for i in range(0, len(documents), batch_size):
    batch = documents[i:i + batch_size]
    task = index.add_documents(batch)
    index.wait_for_task(task.task_uid)

{
  "hits": [
    {
      "id": 1234,
      "title": "Sony WH-1000XM5 wireless headphones",
      "_formatted": {
        "title": "Sony WH-1000XM5 wireless <mark>headphones</mark>"
      }
    }
  ],
  "query": "headphon es sony",
  "processingTimeMs": 4,
  "totalHits": 38,
  "page": 1,
  "hitsPerPage": 20
}

{
  "query": {
    "bool": {
      "should": [
        {
          "multi_match": {
            "query": "наушниик",
            "fields": ["title^3", "brand^2", "description"],
            "fuzziness": "AUTO",
            "prefix_length": 2,
            "max_expansions": 50
          }
        },
        {
          "match_phrase": {
            "title": {
              "query": "наушниик",
              "slop": 2
            }
          }
        }
      ]
    }
  }
}