Getting Started
This guide will help you get started with the Mindee Python OCR SDK to easily extract data from your documents.
The Python OCR SDK supports invoice, passport, receipt OCR APIs and custom-built API from the API Builder.
You can view the source code on GitHub, and the package on PyPI.
Prerequisite
- Download and install Python. This library is officially supported on Python
3.7to3.11. Note: support for3.12is on its way, but currently untested. - Download and install pip package manager.
Installation
To quickly get started with the Python OCR SDK anywhere, the preferred installation method is via pip.
pip install mindee
Development Installation
If you'll be modifying the source code, you'll need to install the development requirements to get started.
- First clone the repo.
git clone [email protected]:mindee/mindee-api-python.git
- Then navigate to the cloned directory and install all development requirements.
cd mindee-api-python
pip install -e ".[dev,test]"
Updating the Version
It is important to always check the version of the Mindee OCR SDK you are using, as new and updated features won’t work on old versions.
To check the installed version:
pip show mindee
To get the latest version:
pip install mindee --upgrade
To install a specific version:
pip install mindee==<your_version>
Usage
To get started with Mindee's APIs, you need to create a Client and you're ready to go.
Let's take a deep dive into how this works.
Initializing the Client
The Client centralizes document configurations in a single object.
The Client requires your API key.
You can either pass these directly to the constructor or through environment variables.
Pass the API key directly
from mindee import Client
# Init with your API key
mindee_client = Client(api_key="my-api-key")
Set the API key in the environment
API keys should be set as environment variables, especially for any production deployment.
The following environment variable will set the global API key:
MINDEE_API_KEY="my-api-key"
Then in your code:
from mindee import Client
# Init without an API key
mindee_client = Client()
Setting the Request Timeout
The request timeout can be set using an environment variable:
MINDEE_REQUEST_TIMEOUT=200
Loading a Document File
Before being able to send a document to the API, it must first be loaded.
You don't need to worry about different MIME types, the library will take care of handling
all supported types automatically.
Once a document is loaded, interacting with it is done in exactly the same way, regardless
of how it was loaded.
There are a few different ways of loading a document file, depending on your use case:
Path
Load from a file directly from disk. Requires an absolute path, as a string.
input_doc = mindee_client.source_from_path("/path/to/the/invoice.pdf")
File Object
A normal Python file object with a path. Must be in binary mode.
with open("/path/to/the/receipt.jpg", 'rb') as fo:
input_doc = mindee_client.source_from_file(fo)
Base64
Requires a base64 encoded string.
Note: The original filename is required when calling the method.
b64_string = "/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAgGBgcGBQgHBwcJCQgKDBQNDAsLD...."
input_doc = mindee_client.source_from_b64string(b64_string, "receipt.jpg")
Bytes
Requires raw bytes.
Note: The original filename is required when calling the method.
raw_bytes = b"%PDF-1.3\n%\xbf\xf7\xa2\xfe\n1 0 ob..."
input_doc = mindee_client.source_from_bytes(raw_bytes, "invoice.pdf")
Loading from bytes is useful when using FastAPI UploadFile objects.
@app.post("/process-file")
async def upload(upload: UploadFile):
input_doc = mindee_client.source_from_bytes(
upload.file.read(),
filename=upload.filename
)
URL
Allows sending an URL directly.
Note: No local operations can be performed on the input (such as removing pages from a PDF).
input_doc = mindee_client.source_from_url(url="https://www.example.com/invoice.pdf")
Sending a File
To send a file to the API, we need to specify how to process the document.
This will determine which API endpoint is used and how the API return will be handled internally by the library.
More specifically, we need to set a mindee.product class as the first parameter of the parse method.
This is because the parse method's' return type depends on its first argument.
Product classes inherit from the base mindee.parsing.common.inference class.
More information is available in each document-specific guide.
Off-the-Shelf Documents
Simply setting the correct class and passing the input document is enough:
result = mindee_client.parse(product.InvoiceV4, input_doc)
Custom Documents
The endpoint to use must be created beforehands and subsequently passed to the endpoint argument of the parse method:
custom_endpoint = mindee_client.create_endpoint(
"my-endpoin-url",
"my-account-name",
# "my-version" # optional
)
result = mindee_client.parse(product.CustomV1, input_doc, endpoint=custom_endpoint)
This is because the CustomV1 class is enough to handle the return processing, but the actual endpoint needs to be specified.
Processing the Response
Results of a prediction can be retrieved in two different places:
Document Level Prediction
The document attribute is an object specific to the type of document being processed.
It is an instance of the Document class, to which a generic type is given.
It contains the data extracted from the entire document, all pages combined.
It's possible to have the same field in various pages, but at the document level only the highest confidence field data will be shown (this is all done automatically at the API level).
Usage:
print(resp.document)
A document's fields (attributes) can be accessed through it's prediction attribute, which have types that can vary from one product to another.
These attributes are detailed in each product's respective guide.
Page Level Prediction
The pages attribute is a list of Page objects. Page is a wrapper around elements that extend the Document class.
The prediction of a Page inherits from the product's own Document, and adds all page-specific fields to it.
The order of the elements in the list matches the order of the pages in the document.
All response objects have a pages property, regardless of the number of pages.
Single-page documents will have a single entry.
Iteration over pages is done like with any list, for example:
for page in resp.pages:
print(page)
Questions?
Updated 2 months ago