Search on your site doesn't find "ноутбук" when a user types "лэптоп"? Or "смартфон" returns no results for "телефон"? This is a typical problem of the standard Elasticsearch analyzer: it doesn't understand synonyms and language morphology. To make search work properly, we develop custom analyzers turnkey. Over 5 years, we've tuned search for 50+ projects, including e-commerce stores and news portals. A custom analyzer consists of a tokenizer and token filters. They transform text into a set of tokens for the index. In this article, we'll break down how to assemble such an analyzer for Russian, configure synonyms and autocomplete, and look at typical pitfalls and how to avoid them.
Up to 70% of users don't find the right product due to incorrect analyzer configuration. This reduces conversion by 20–30%. Configuring a custom analyzer solves the problem by selecting filters tailored to your domain.
| Filter | Purpose | Example token |
|---|---|---|
lowercase |
Lowercasing | Ноутбук → ноутбук |
stop |
Stop word removal | на, в, с |
stemmer |
Stemming | ноутбуки → ноутбук |
synonym |
Synonyms | ноутбук ↔ лэптоп |
edge_ngram |
N-grams for autocomplete | ноут → н, но, ноу, ноут |
Why a custom analyzer beats the built-in?
The built-in russian analyzer uses the Snowball algorithm for stemming. For most tasks, that's enough, but if you need precise morphology or a custom set of stop words, we create a custom analyzer. It gives you full control over the filter chain: you decide which stop words to remove, how to handle synonyms, and whether stemming is needed. Additionally, a custom configuration allows adding transliteration for searching by translit (e.g., "noutbuk" → "ноутбук").
How to set up a language analyzer for Russian?
Basic configuration
Example with custom stop words and a field for exact match:
{ "settings": { "analysis": { "filter": { "russian_stop": { "type": "stop", "stopwords": "_russian_" }, "russian_stemmer": { "type": "stemmer", "language": "russian" }, "custom_stopwords": { "type": "stop", "stopwords": ["это", "также", "при", "для", "что", "как"] } }, "analyzer": { "russian_custom": { "type": "custom", "tokenizer": "standard", "filter": ["lowercase", "russian_stop", "custom_stopwords", "russian_stemmer"] }, "russian_exact": { "type": "custom", "tokenizer": "standard", "filter": ["lowercase"] } } } } } Adding transliteration
Users often type translit: "noutbuk" instead of "ноутбук". Transliteration is implemented via a char_filter of type mapping, where each Cyrillic letter is mapped to a Latin equivalent. The full mapping list is in the documentation — we include it in the analyzer. For advanced transliteration, we use the analysis-icu plugin with ICU transformers, which correctly handles Unicode normalization.
How to set up synonyms without reindexing?
Synonyms allow finding documents with different wording. There are two approaches:
| Parameter | Index-time synonyms | Search-time synonyms |
|---|---|---|
| Update | Requires reindexing | Without reindexing, via _reload_search_analyzers |
| Search performance | Faster (fewer tokens in query) | Slower (additional processing) |
| Flexibility | Low | High |
The recommended option is search-time synonyms. They allow changing rules without reindexing, which is critical for e-commerce stores with rapidly changing assortments. For example, a synonym file config/synonyms/synonyms.txt:
ноутбук, лэптоп, laptop, notebook смартфон, телефон, мобильный телевизор, тв, tv Analyzer configuration with synonyms (only for search):
{ "settings": { "analysis": { "filter": { "synonym_filter": { "type": "synonym", "synonyms_path": "synonyms/synonyms.txt", "updateable": true } }, "analyzer": { "search_analyzer": { "type": "custom", "tokenizer": "standard", "filter": ["lowercase", "russian_stop", "synonym_filter"] }, "index_analyzer": { "type": "custom", "tokenizer": "standard", "filter": ["lowercase", "russian_stop", "russian_stemmer"] } } } } } The flag "updateable": true is mandatory for search-only synonyms. It allows reloading synonyms via _reload_search_analyzers without reindexing. Budget saving: you don't need to rebuild the index every time the dictionary changes.
Setting up autocomplete with Edge N-gram
For autocomplete during input, you need an analyzer with edge_ngram:
"filter": { "edge_ngram_filter": { "type": "edge_ngram", "min_gram": 2, "max_gram": 15 } } At search time, use an analyzer without N-gram (e.g., just lowercase), otherwise a query for "ноут" will match all N-gram tokens, causing false positives.
Multilingual indexes
Two approaches: one index with fields for each language or a separate index per language. For multisite setups, one index is more convenient:
"properties": { "title_ru": { "type": "text", "analyzer": "russian_custom" }, "title_en": { "type": "text", "analyzer": "english" }, "title_de": { "type": "text", "analyzer": "german" } } Query across multiple languages:
{ "query": { "multi_match": { "query": "search term", "fields": ["title_ru^2", "title_en", "title_de"] } } } Boost ^2 for the Russian field if your primary audience is Russian-speaking.
What's included in development and timelines
We provide a fully turnkey solution. Development includes:
- Analysis of your data and search requirements
- Designing a filter chain for your tasks
- Configuring synonyms with online update capability
- Integration with your existing index
- Testing on real queries (at least 100 test scenarios)
- Documentation on settings and maintenance
- Training for your team (up to 2 hours online)
Development and testing of a custom analyzer for one language — from 1 business day. Configuring synonyms with a file and update mechanism — half a day more. Multilingual configuration with 3–5 languages — 2–3 days, including search relevance tests. Cost is calculated individually.
Order analyzer configuration for your project — we'll find the optimal setup.
How we approach the setup
- Current search audit — collect query logs, identify problematic scenarios.
- Analyzer design — define filter set, synonyms, stop words.
- Prototyping — deploy a test index and check on real data.
- Testing — use the
_analyzeAPI to verify each filter. For example:
POST /products/_analyze { "analyzer": "russian_custom", "text": "Ноутбуки и лэптопы для работы" } The response contains a list of tokens — their count and type should match expectations (stemming trims endings, synonyms expand tokens). 5. Optimization — adjust settings based on test results. 6. Deployment — update the production index without downtime. 7. Monitoring — after launch, track search metrics (CTR, empty results).
Why entrust configuration to professionals?
Incorrect analyzer configuration can reduce search relevance by 2–3 times. Our certified engineers guarantee that after setup, search will be accurate and fast. Get a consultation on analyzer development — we'll evaluate your project and suggest the best solution. Contact us to discuss details.
Learn more about Elasticsearch analyzers.