Text

field_match

An extractor which can match a field from ranking event over an item field. In practice, it can be useful in search related tasks, when you need to match a search query over multiple separate fields in document, like title-tags-category.

Field match extractor supports the following matching methods:

• BM25: a Lucene-specific BM25 score between ranking and item fields (for example, between query and item title)

• ngram: split item/query fields to N-grams and compute intersection over union score

• term: use Lucene to perform language specific tokenization

• bert: build LLM embeddings for item/query fields and compute a distance between them

Dataset example

Given this metadata event:

{
  "event": "item",
  "id": "81f46c34-a4bb-469c-8708-f8127cd67d27",
  "item": "item1", 
  "timestamp": "1599391467000", 
  "fields": [
    {"name": "title", "value": "red socks"},
    {"name": "category", "value": "socks"},
    {"name": "brand", "value": "buffalo"},
    {"name": "description", "value": "lorem ipsum dolores sit amet"}
  ]
}

And a following ranking event:

{
  "event": "ranking",
  "id": "81f46c34-a4bb-469c-8708-f8127cd67d27",
  "timestamp": "1599391467000",
  "user": "user1",
  "session": "session1",
  "fields": [
      {"name": "query", "value": "sock"}
  ],
  "items": [
    {"id": "item3"},
    {"id": "item1"},
    {"id": "item2"} 
  ]
}

BM25 score

As BM25 formula requires term frequencies and some other index statistics, using BM25 requires you to build the term-freq dictionary beforehead, see the CLI termfreq docs on how to do it.

Having the term-freq.json file in hand, you can then configure Metarank to compute BM25 score between ranking field (for example, query) and item field (like title):

  - name: title_match
    type: field_match
    rankingField: ranking.query
    itemField: item.title
    method:
      type: bm25
      language: english
      termFreq: "/path/to/term-freq.json"

Ngram matching

With the following config file snippet you can do a per-field matching of ranking.query field over item.title field of the items in the ranking with 3-grams:

- name: title_match
  type: field_match
  itemField: item.title // must be a string
  rankingField: ranking.query // must be a string
  method:
    type: ngram // for now only ngram and term are supported
    language: en // ISO-639-1 language code
    n: 3
  refresh: 0s // optional, how frequently we should update the value, 0s by default
  ttl: 90d // optional, how long should we store this field

Term matching

In a similar way you can do the same with term matching:

- name: title_match
  type: field_match
  itemField: item.title // must be a string
  rankingField: ranking.query // must be a string
  method:
    type: term // for now only ngram and term are supported
    language: en // ISO-639-1 language code

Both term and ngram matching methods leverage Lucene for text analysis and support the following set of languages:

• generic: no language specific transformations

• en: English

• cz: Czech

• da: Danish

• nl: Dutch

• et: Estonian

• fi: Finnish

• fr: French

• de: German

• gr: Greek

• it: Italian

• no: Norwegian

• pl: Polish

• pt: Portuguese

• es: Spanish

• sv: Swedish

• tr: Turkish

• ar: Arabic

• zh: Chinese

• ja: Japanese

Both term and ngram method share the same approach to the text analysis:

• text line is split into terms (using language-specific method)

• stopwords are removed

• for non-generic languages each term is stemmed

• then terms/ngrams from item and ranking are scored using intersection/union method.

LLM Bi-Encoders

This text matching method:

• computes an embedding for both query and document

• then computes a cosine between both embeddings.

Semantically-similar query-document pairs will have higher score than irrelevant ones.

Then with the following config snippet we can compute a cosine distance between title and query embeddings:

- type: field_match
  name: title_query_match
  rankingField: ranking.query
  itemField: item.title
  distance: cos # optional, default cos, options: cos/dot 
  method:
    type: bi-encoder
    model: metarank/all-MiniLM-L6-v2 # optional, can be only cache-based
    dim: 384 # required, dimensionality of the embedding
    itemFieldCache: /path/to/item.embedding # optional, pre-computed embedding cache for items 
    rankingFieldCache: /path/to/query.embedding # optional, pre-computed embedding cache for rankings

Metarank supports two embedding methods:

• bi-encoder: ONNX-encoded versions of the sentence-transformers models. See the metarank HuggingFace namespace for a list of currently supported models.
• cross-encoder: ONNX-encoded versions of sentence-transformers

For both models, Metarank supports fetching model directly from the HuggingFace Hub, or loading it from a local dir, depending on the model handle format:

• namespace/model: fetch model from the HFHub

• file:///<path>/<to>/<model dir>: load ONNX-encoded embedding model from a local file.

Using CSV cache of precomputed embeddings

In some performance-sensitive cases you don't want to compute embeddings in realtime, but only use offline precomputed ones. This is possible with the csv field_match method:

- type: field_match
  name: title_query_match
  rankingField: ranking.query
  itemField: item.title
  distance: cos # optional, default cos, options: cos/dot 
  method:
    type: bi-encoder # note that there is no model reference, only caches
    dim: 384
    itemFieldCache: /path/to/item.embedding
    rankingFieldCache: /path/to/query.embedding

In this case Metarank will load item and query embeddings from a CSV file in the following format:

itemFieldCache:

item1,0,1,2,3,4,5
item2,5,4,3,2,1,9
item3,1,1,1,1,1,1

rankingFieldCache:

bananas,0,1,2,3,4,5
red socks,5,4,3,2,1,9
stone,1,1,1,1,1,1

• when both query and item embeddings are present, then field_match will produce a cosine distance between them.

• when at least one of the embeddings is missing, then field_match with csv method will produce a nil missing value.

• when at least one of the embeddings is missing, then field_match with transformer method will compute the embedding real-time.

LLM Cross-encoders

Cross-encoders are quite similar to bi-encoders, but instead of separately computing embedding for query and document, they feed both texts to the neural network, which produces the matching score.

Compared to the bi-encoder approach:

• cross-encoders are much more precise, even generic pre-trained models.

• require much more resources, as there's no way to pre-compute embeddings for docs and queries - you need to perform a full neural inference query-time.

Enabling cross-encoders in Metarank can be done with the following snippet:

- type: field_match
  name: title_query_match
  rankingField: ranking.query
  itemField: item.title
  method:
    type: cross-encoder
    model: metarank/ce-msmarco-MiniLM-L6-v2 # optional, can be only cache-based
    cache: /path/to/ce.cache # optional, pre-computed query-doc scores

Note that as cross-encoders are very CPU heavy to run, you can pre-compute a set of query-doc scores offline and supply Metarank with a cache in the following CSV format:

query1,doc1,0.7
query1,doc2,0.1
query2,doc3,0.2