Similarity

Similarity

The Similarity pipeline computes similarity between queries and list of text using a text classifier.

This pipeline supports both standard text classification models and zero-shot classification models. The pipeline uses the queries as labels for the input text. The results are transposed to get scores per query/label vs scores per input text.

Cross-encoder models are supported via the crossencode=True constructor parameter. Late interaction (i.e. ColBERT) models are also supported via the lateencode=True constructor parameter. CrossEncoder and LateEncoder pipelines back each of these models and can be instantiated directly as well.

Example

The following shows a simple example using this pipeline.

from txtai.pipeline import Similarity

## Create and run pipeline
similarity = Similarity()
similarity("feel good story", [
    "Maine man wins $1M from $25 lottery ticket",
    "Don't sacrifice slower friends in a bear attack"
])

See the link below for a more detailed example.

Notebook| Description|
---|---|---
Add semantic search to Elasticsearch| Add semantic search to existing search systems|

Configuration-driven example

Pipelines are run with Python or configuration. Pipelines can be instantiated in configuration using the lower case name of the pipeline. Configuration-driven pipelines are run with workflows or the API.

config.yml

## Create pipeline using lower case class name
similarity:

Run with Workflows

from txtai import Application

## Create and run pipeline with workflow
app = Application("config.yml")
app.similarity("feel good story", [
    "Maine man wins $1M from $25 lottery ticket",
    "Don't sacrifice slower friends in a bear attack"
])

Run with API

CONFIG=config.yml uvicorn "txtai.api:app" &

curl \
  -X POST "http://localhost:8000/similarity" \
  -H "Content-Type: application/json" \
  -d '{"query": "feel good story", "texts": ["Maine man wins $1M from $25 lottery ticket", "Dont sacrifice slower friends in a bear attack"]}'

Methods

Python documentation for the pipeline.

__init__(path=None, quantize=False, gpu=True, model=None, dynamic=True, crossencode=False, lateencode=False, **kwargs)

Source code in txtai/pipeline/text/similarity.py

17
18
19
20
21
22
23
24
25
26
27
28
``` |  ```
def init(self, path=None, quantize=False, gpu=True, model=None, dynamic=True, crossencode=False, lateencode=False, kwargs):
    self.crossencoder, self.lateencoder = None, None

    if lateencode:
        # Load a late interaction encoder if lateencode set to True
        self.lateencoder = LateEncoder(path=path, gpu=gpu, kwargs)
    else:
        # Use zero-shot classification if dynamic is True and crossencode is False, otherwise use standard text classification
        super().init(path, quantize, gpu, model, False if crossencode else dynamic, **kwargs)

        # Load as a cross-encoder if crossencode set to True
        self.crossencoder = CrossEncoder(model=self.pipeline) if crossencode else None

---|---

__call__(query, texts, multilabel=True, **kwargs)

Computes the similarity between query and list of text. Returns a list of (id, score) sorted by highest score, where id is the index in texts.

This method supports query as a string or a list. If the input is a string, the return type is a 1D list of (id, score). If text is a list, a 2D list of (id, score) is returned with a row per string.

Parameters:

Returns:

Source code in txtai/pipeline/text/similarity.py

31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
``` |  ```
def call(self, query, texts, multilabel=True, kwargs):
    “””
    Computes the similarity between query and list of text. Returns a list of
    (id, score) sorted by highest score, where id is the index in texts.

    This method supports query as a string or a list. If the input is a string,
    the return type is a 1D list of (id, score). If text is a list, a 2D list
    of (id, score) is returned with a row per string.

    Args:
        query: query text|list
        texts: list of text
        multilabel: labels are independent if True, scores are normalized to sum to 1 per text item if False, raw scores returned if None
        kwargs: additional keyword args

    Returns:
        list of (id, score)
    “””

    if self.crossencoder:
        # pylint: disable=E1102
        return self.crossencoder(query, texts, multilabel)

    if self.lateencoder:
        return self.lateencoder(query, texts)

    # Call Labels pipeline for texts using input query as the candidate label
    scores = super().call(texts, [query] if isinstance(query, str) else query, multilabel, kwargs)

    # Sort on query index id
    scores = [[score for _, score in sorted(row)] for row in scores]

    # Transpose axes to get a list of text scores for each query
    scores = np.array(scores).T.tolist()

    # Build list of (id, score) per query sorted by highest score
    scores = [sorted(enumerate(row), key=lambda x: x[1], reverse=True) for row in scores]

    return scores[0] if isinstance(query, str) else scores

---|---