Tokenizer vs token filters

I want to publish some detailed use cases.

Edge n-gram tokenizer (default)

By default this tokenizer treats all text as a single token because by default token can contain any characters (including spaces).

GET {ELASTICSEARCH_URL}/_analyze

{
  "tokenizer": "edge_ngram",
  "text": "How are you?"
}

Result:

["H", "Ho"]

Explanation: one token, min_gram = 1, max_gram = 2.

Edge n-gram tokenizer (custom without token_chars)

PUT {ELASTICSEARCH_URL}/custom_edge_ngram

{
  "settings": {
    "analysis": {
      "analyzer": {
        "custom_edge_ngram": {
          "tokenizer": "custom_edge_ngram_tokenizer"
        }
      },
      "tokenizer": {
        "custom_edge_ngram_tokenizer": {
          "type": "edge_ngram",
          "min_gram": 2,
          "max_gram": 7
        }
      }
    }
  }
}

GET {ELASTICSEARCH_URL}/custom_edge_ngram/_analyze

{
  "analyzer": "custom_edge_ngram",
  "text": "How old are you?"
}

Result:

["Ho", "How", "How ", "How o", "How ol", "How old"]

Explanation: still one token, min_gram = 2, max_gram = 7.

Edge n-gram tokenizer (custom with token_chars)

PUT {ELASTICSEARCH_URL}/custom_edge_ngram_2

{
  "settings": {
    "analysis": {
      "analyzer": {
        "custom_edge_ngram": {
          "tokenizer": "custom_edge_ngram_tokenizer"
        }
      },
      "tokenizer": {
        "custom_edge_ngram_tokenizer": {
          "type": "edge_ngram",
          "min_gram": 2,
          "max_gram": 7,
          "token_chars": ["letter"]
        }
      }
    }
  }
}

GET {ELASTICSEARCH_URL}/custom_edge_ngram_2/_analyze

{
  "analyzer": "custom_edge_ngram",
  "text": "How old are you?"
}

Result:

["Ho", "How", "ol", "old", "ar", "are", "yo", "you"]

Explanation: 4 tokens How, old, are, you (tokens can contain only letters because of token_chars), min_gram = 2, max_gram = 7, but max token length in the sentence is 3.

Edge n-gram token filter

Tokenizer converts text to stream of tokens.

Token filter works with each token of the stream.

Token filter can modify stream by adding, updating, deleting its tokens.

Let's use standard tokenizer.

GET {ELASTICSEARCH_URL}/_analyze

{
  "tokenizer": "standard",
  "text": "How old are you?"
}

Result:

["How", "old", "are", "you"]

Now let's add token filter.

GET {ELASTICSEARCH_URL}/_analyze

{
  "tokenizer": "standard",
  "filter": [
    { "type": "edge_ngram",
      "min_gram": 2,
      "max_gram": 7
    }
  ],
  "text": "How old are you?"
}

Result:

["Ho", "How", "ol", "old", "ar", "are", "yo", "you"]

Explanation: edge_nram for each token.

A tokenizer will split the whole input into tokens and a token filter will apply some transformation on each token.

For instance, let's say the input is The quick brown fox. If you use an edgeNGram tokenizer, you'll get the following tokens:

• T
• Th
• The
• The (last character is a space)
• The q
• The qu
• The qui
• The quic
• The quick
• The quick (last character is a space)
• The quick b
• The quick br
• The quick bro
• The quick brow
• The quick brown
• The quick brown (last character is a space)
• The quick brown f
• The quick brown fo
• The quick brown fox

However, if you use a standard tokenizer which will split the input into words/tokens, and then an edgeNGram token filter, you'll get the following tokens

• T, Th, The
• q, qu, qui, quic, quick
• b, br, bro, brow, brown
• f, fo, fox

As you can see, choosing between an edgeNgram tokenizer or token filter depends on how you want to slice and dice your text and how you want to search it.

I suggest having a look at the excellent elyzer tool which provides a way to visualize the analysis process and see what is being produced during each step (tokenizing and token filtering).

As of ES 2.2, the _analyze endpoint also supports an explain feature which shows the details during each step of the analysis process.