Skip to main content

Edit this page

nearest_neighbors

The nearest_neighbors template uses a pre-trained embedding model to expand each input token with the terms whose vectors lie closest to it. In effect it enriches the text with semantically related words, so a document indexed through this template can be found by synonyms and near-synonyms it never literally contained — a recall-oriented complement to exact full-text matching.

How it works​

For every token in the input the analyzer asks a fastText embedding model loaded from modellocation for its topk nearest neighbors and emits those neighbor words as additional terms. Applied at index time it broadens what a document can match; applied to the query it broadens what the query reaches. For example a cooking model might expand "cake" into related terms such as three-tiered and wham.

The model file is required and must be reachable from the server process at the path given in modellocation; the dictionary cannot be created without a loadable model. Where classification tags a document with predicted category labels, nearest_neighbors instead grows its vocabulary with related terms.

Options​

OptionTypeDefaultDescription
modellocationstringrequiredPath to the fastText model file, reachable from the server
topkinteger1Number of nearest neighbors to emit per input token

Usage​

Point modellocation at a trained fastText model and choose how many neighbors to add per token:

Query
CREATE TEXT SEARCH DICTIONARY nn_dict (    template = 'nearest_neighbors',    MODELLOCATION = '/models/cooking.bin',    TOPK = 2);

Attached to a text column in a USING inverted index, the dictionary indexes each document under both its own tokens and their nearest neighbors, widening recall. The emitted neighbors depend on the model; the example above uses a small cooking model.

See also​