Elasticsearch is all the buzzwords recently. There are projects using it to fast distributed search, as a document database and as a tool for data analysis. I've wanted to check out Elasticsearch for some time already, even more, since I've been at lecture about aggregations in Elasticsearch by Honza Král at PyConPL 2013 . I recently got a chance to check it out migrating project from whoosh to elasticsearch. And first thing I noticed about Elasticsearch from solutions I've been using so far are analyzers.

Unlike Sphinx or Whoosh, where you determine how to analyze a field by search engine, in Elasticsearch, you define an analyzer for a field to analyze both the field, and query for the field as well.

{
    "mappings": {
        "document": {
            "properties": {
                "title": {
                    "type": "string",
                    "analyzer": "autocomplete_analyzer",
                }
            }
        }
    }
}

This mapping definition will cause Elasticsearch to index title field according to autocomplete_analyzer and query as well. That's fine for most cases, but eg. for autocomplete usually tries to suggest where user can go further from what he typed, not what else he might have typed in.

Elasticsearch provides a possibility to define separate analyzer index_analyzer for index and search_analyzer for search query.

{
    "mappings": {
        "document": {
            "properties": {
                "title": {
                    "type": "string",
                    "index_analyzer": "autocomplete_analyzer",
                    "search_analyzer": "simple"
                }
            }
        }
    }
}

This will allow to limit suggestions to take account whole words user has already typed in. There's one caveat though. If you leave also the default analyzer definition, elasticsearch won't take into account index_analyzer and search_analyzer, but will create an additional search_query_analyzer. And this, at least for me, haven't returned the results I expected.

Of course, if you standard analyzers won't be what you seek, there's always an option to define custom analyzers, custom tokenizers, and custom filters.