Konfuzio SDK

The Konfuzio Software Development Kit (Konfuzio SDK) provides a Python API to interact with the Konfuzio Server.

Features

The SDK allows you to retrieve visual and text features to build your own document models. Konfuzio Server serves as an UI to define the data structure, manage training/test data and to deploy your models as API.

* Under fair use policy: We will impose 10 pages/hour throttling eventually.

Installation

As developer register on our public HOST for free: https://app.konfuzio.com

Then you can use pip to install Konfuzio SDK and run init:

pip install konfuzio_sdk

konfuzio_sdk init

The init will create a Token to connect to the Konfuzio Server. This will create variables KONFUZIO_USER, KONFUZIO_TOKEN and KONFUZIO_HOST in an .env file in your working directory.

Find the full installation guide here or setup PyCharm as described here.

Basics

from konfuzio_sdk.data import Project, Document

# Initialize the project:
my_project = Project(id_='YOUR_PROJECT_ID')

# Get any project online
doc: Document = my_project.get_document_by_id('DOCUMENT_ID_ONLNIE')

# Get the Annotations in a Document
doc.annotations()

# Filter Annotations by Label
label = my_project.get_label_by_name('MY_OWN_LABEL_NAME')
doc.annotations(label=label)

# Or get all Annotations that belong to one Label
label.annotations

# Force a project update. To save time documents will only be updated if they have changed.
my_project.update()

Find more explanations in the Examples.

Regex

Pro Tip: Read our technical blog post Automated Regex to find out how we use Regex to detect outliers in our annotated data.

from konfuzio_sdk.regex import suggest_regex_for_string
from konfuzio_sdk.data import Project, Label

my_project = Project(id_='YOUR_PROJECT_ID')
label: Label = my_project.get_label_by_name('MY_OWN_LABEL_NAME')

# Get Regex tokens to capture (nearly) all annotations of this Label
tokens = label.tokens()
assert tokens == [
    "(?P<GesamtBrutto_N_4420363_2111>\\d\\.\\d\\d\\d\\,\\d\\d)",
    "(?P<GesamtBrutto_N_9812334_1498>\\d\\d\\d\\,\\d\\d)"
]

# Get optimize regex (Can be multiple if multiple create a higher accuracy than a single one)
label_regex = label.regex()
assert label_regex == [
    "[ ]+(?:(?P<GesamtBrutto_N_4420363_2111>\\d\\.\\d\\d\\d\\,\\d\\d)|(?P<GesamtBrutto_N_9812334_1498>\\d\\d\\d\\,\\d\\d))\n"
]

# Suggest a RegEx for a string without optimization
regex = suggest_regex_for_string('Date: 20.05.2022')
assert regex == 'Date:[ ]+\\d\\d\\.\\d\\d\\.\\d\\d\\d\\d'

Tokenizer

Create a Tokenizer based on a Regex and evaluate it on a Document level.

from konfuzio_sdk.data import Project
from konfuzio_sdk.tokenizer.regex import RegexTokenizer

my_project = Project(id_='YOUR_PROJECT_ID')
document = my_project.get_document_by_id(document_id='YOUR_DOCUMENT_ID')

# Define the Regex expression
regex = r'[^ \n\t\f]+'

# Build a Tokenizer based on Regex 
tokenizer = RegexTokenizer(regex=regex)
assert tokenizer.regex == regex

# Evaluate the Tokenizer in a Document
evaluation = tokenizer.evaluate(document)

# Ratio of correct Spans found by the Tokenizer in the Document
ratio_of_spans_found = evaluation.is_found_by_tokenizer.sum() / evaluation.is_correct.sum()

Add visual features to text

Calculate the bounding box of a Span using the start and end character.

from pprint import pprint

from konfuzio_sdk.data import Project, LabelSet, Label, AnnotationSet, Annotation, Span
import os

OFFLINE_PROJECT = os.path.join("tests", "example_project_data")

my_project = Project(id_=None, project_folder=OFFLINE_PROJECT)  # use offline data and don't connect to Server
document = my_project.get_document_by_id(44823)
label = Label(project=my_project)
label_set = LabelSet(project=my_project, categories=[document.category])
annotation_set = AnnotationSet(document=document, label_set=label_set)

span = Span(start_offset=60, end_offset=65)

annotation = Annotation(
    label=label,
    annotation_set=annotation_set,
    label_set=label_set,
    document=document,
    spans=[span],
)

span_with_bbox_information = span.bbox()

pprint(span_with_bbox_information.__dict__)

{'_line_index': None,
 '_page_index': 0,
 'annotation': Annotation (None) None (60, 65),
 'bottom': 32.849,
 'end_offset': 65,
 'id_local': 74,
 'start_offset': 60,
 'top': 23.849,
 'x0': 426.0,
 'x1': 442.8,
 'y0': 808.831,
 'y1': 817.831}

CLI

We provide the basic function to create a new Project via CLI:

konfuzio_sdk create_project YOUR_PROJECT_NAME

You will see "Project {YOUR_PROJECT_NAME} (ID {YOUR_PROJECT_ID}) was created successfully!" printed.

And download any project via the id:

konfuzio_sdk download_data YOUR_PROJECT_ID

Tutorials

• Train a Konfuzio SDK Model to Extract Information From Payslip Documents : An example of how Konfuzio SDK package can be used in a pipeline to have an easy feedback workflow can be seen in this tutorial
• Automate Annotations with Regex : An example of how to create regex-based annotations in a Konfuzio project.
• Retrain Flair NER-Ontonotes-Fast with Human Revised Annotations : An example of how Konfuzio SDK package can be used in a pipeline to have an easy feedback workflow can be seen in this tutorial
• Count Relevant Expressions in Annual Reports : An example of how to retrieve structured and organized information from documents.

The Konfuzio Server Tutorial:

References

• Konfuzio SDK Python API - Source Code
• Konfuzio Server REST API
• How to Contribute
• Issue Tracker
• MIT License
• Konfuzio Homepage

Supported CRUD Operations