Suggesters & autocomplete

Suggesters return word- or phrase-level suggestions for a query string — “did you mean?” corrections and type-ahead completion — served from purpose-built data structures rather than the main inverted-index.

Why it matters

Autocomplete and spell-correction must respond in single-digit milliseconds as the user types each keystroke, far faster than a match/wildcard query against the full index can deliver. Doing type-ahead with leading-wildcard or prefix queries scans the whole term dictionary per shard and collapses under load; suggesters move that cost to index time.

How it works

Elasticsearch ships four suggesters, each backed by a different structure.

Suggester	Purpose	Backing structure
term	per-token spell correction	edit-distance over terms
phrase	whole-phrase “did you mean”	n-gram language model
completion	prefix type-ahead	in-memory FST
context	filtered/boosted completion	FST + category/geo context

• completion type — declare a field as "type": "completion"; it builds a Finite State Transducer (tries-like) held in heap, giving prefix lookups in roughly the length of the prefix.
• Weights & dedup — each input carries a weight; suggestions return ranked by weight, not BM25.
• context suggester — attach category or geo contexts so “piz” suggests only menu items at this restaurant.
• fuzzy on completion — typo-tolerant prefixes via fuzziness, with prefix_length to anchor the start.

Example

PUT /place { "mappings": { "properties": {
  "name": { "type": "completion" } } } }
PUT /place/_doc/1 { "name": { "input": ["Pizza Hut","Pizza"], "weight": 30 } }

POST /place/_search { "suggest": { "s": {
  "prefix": "piz", "completion": { "field": "name", "fuzzy": { "fuzziness": 1 } } } } }

Typing piz returns “Pizza”/“Pizza Hut” from the FST; pza still hits via fuzziness 1.

Pitfalls

• Heap cost — the completion FST is loaded entirely into memory; millions of distinct inputs inflate heap and can pressure the node.
• No mid-word match — completion only matches from the start; for infix (“auto” in “kitautoware”) use an edge_ngram analyzer on a normal text field instead.
• Reindex to update — changing the completion field or its inputs needs a reindex; you can’t patch the FST in place.
• term/phrase need a real corpus — they suggest from indexed terms, so a sparse index yields poor corrections.

See also

• fuzzy-wildcard-search
• tries