Amazon Textract

Amazon Textract is a machine learning (ML) service that automatically extracts text, handwriting, and data from scanned documents.

It goes beyond simple optical character recognition (OCR) to identify, understand, and extract data from forms and tables. Today, many companies manually extract data from scanned documents such as PDFs, images, tables, and forms, or through simple OCR software that requires manual configuration (which often must be updated when the form changes). To overcome these manual and expensive processes, Textract uses ML to read and process any type of document, accurately extracting text, handwriting, tables, and other data with no manual effort.

Textract supports JPEG, PNG, PDF, and TIFF file formats; more information is available in the documentation.

The following examples demonstrate the use of Amazon Textract in combination with LangChain as a DocumentLoader.

%pip install -qU  boto3 langchain-openai tiktoken python-dotenv

%pip install -qU  "amazon-textract-caller>=0.2.0"

Example 1: Loading from a local file

The first example uses a local file, which internally will be sent to Amazon Textract sync API DetectDocumentText.

Local files or URL endpoints like HTTP:// are limited to one page documents for Textract. Multi-page documents have to reside on S3. This sample file is a jpeg.

from langchain_community.document_loaders import AmazonTextractPDFLoader

loader = AmazonTextractPDFLoader("example_data/alejandro_rosalez_sample-small.jpeg")
documents = loader.load()

Output from the file

documents

[Document(page_content='Patient Information First Name: ALEJANDRO Last Name: ROSALEZ Date of Birth: 10/10/1982 Sex: M Marital Status: MARRIED Email Address: Address: 123 ANY STREET City: ANYTOWN State: CA Zip Code: 12345 Phone: [phone] Emergency Contact 1: First Name: CARLOS Last Name: SALAZAR Phone: [phone] Relationship to Patient: BROTHER Emergency Contact 2: First Name: JANE Last Name: DOE Phone: [phone] Relationship FRIEND to Patient: Did you feel fever or feverish lately? Yes No Are you having shortness of breath? Yes No Do you have a cough? Yes No Did you experience loss of taste or smell? Yes No Where you in contact with any confirmed COVID-19 positive patients? Yes No Did you travel in the past 14 days to any regions affected by COVID-19? Yes No Patient Information First Name: ALEJANDRO Last Name: ROSALEZ Date of Birth: 10/10/1982 Sex: M Marital Status: MARRIED Email Address: Address: 123 ANY STREET City: ANYTOWN State: CA Zip Code: 12345 Phone: [phone] Emergency Contact 1: First Name: CARLOS Last Name: SALAZAR Phone: [phone] Relationship to Patient: BROTHER Emergency Contact 2: First Name: JANE Last Name: DOE Phone: [phone] Relationship FRIEND to Patient: Did you feel fever or feverish lately? Yes No Are you having shortness of breath? Yes No Do you have a cough? Yes No Did you experience loss of taste or smell? Yes No Where you in contact with any confirmed COVID-19 positive patients? Yes No Did you travel in the past 14 days to any regions affected by COVID-19? Yes No ', metadata={'source': 'example_data/alejandro_rosalez_sample-small.jpeg', 'page': 1})]

Example 2: Loading from a URL

The next example loads a file from an HTTPS endpoint. It has to be single page, as Amazon Textract requires all multi-page documents to be stored on S3.

from langchain_community.document_loaders import AmazonTextractPDFLoader

loader = AmazonTextractPDFLoader(
    "https://amazon-textract-public-content.s3.us-east-2.amazonaws.com/langchain/alejandro_rosalez_sample_1.jpg"
)
documents = loader.load()

documents

[Document(page_content='Patient Information First Name: ALEJANDRO Last Name: ROSALEZ Date of Birth: 10/10/1982 Sex: M Marital Status: MARRIED Email Address: Address: 123 ANY STREET City: ANYTOWN State: CA Zip Code: 12345 Phone: [phone] Emergency Contact 1: First Name: CARLOS Last Name: SALAZAR Phone: [phone] Relationship to Patient: BROTHER Emergency Contact 2: First Name: JANE Last Name: DOE Phone: [phone] Relationship FRIEND to Patient: Did you feel fever or feverish lately? Yes No Are you having shortness of breath? Yes No Do you have a cough? Yes No Did you experience loss of taste or smell? Yes No Where you in contact with any confirmed COVID-19 positive patients? Yes No Did you travel in the past 14 days to any regions affected by COVID-19? Yes No Patient Information First Name: ALEJANDRO Last Name: ROSALEZ Date of Birth: 10/10/1982 Sex: M Marital Status: MARRIED Email Address: Address: 123 ANY STREET City: ANYTOWN State: CA Zip Code: 12345 Phone: [phone] Emergency Contact 1: First Name: CARLOS Last Name: SALAZAR Phone: [phone] Relationship to Patient: BROTHER Emergency Contact 2: First Name: JANE Last Name: DOE Phone: [phone] Relationship FRIEND to Patient: Did you feel fever or feverish lately? Yes No Are you having shortness of breath? Yes No Do you have a cough? Yes No Did you experience loss of taste or smell? Yes No Where you in contact with any confirmed COVID-19 positive patients? Yes No Did you travel in the past 14 days to any regions affected by COVID-19? Yes No ', metadata={'source': 'example_data/alejandro_rosalez_sample-small.jpeg', 'page': 1})]

Example 3: Loading multi-page PDF documents

Processing a multi-page document requires the document to be on S3. The sample document resides in a bucket in us-east-2 and Textract needs to be called in that same region to be successful, so we set the region_name on the client and pass that in to the loader to ensure Textract is called from us-east-2. You could also to have your notebook running in us-east-2, setting the AWS_DEFAULT_REGION set to us-east-2 or when running in a different environment, pass in a boto3 Textract client with that region name like in the cell below.

import boto3

textract_client = boto3.client("textract", region_name="us-east-2")

file_path = "s3://amazon-textract-public-content/langchain/layout-parser-paper.pdf"
loader = AmazonTextractPDFLoader(file_path, client=textract_client)
documents = loader.load()

Now getting the number of pages to validate the response (printing out the full response would be quite long…). We expect 16 pages.

len(documents)

16

Example 4: Customizing the output format

When Amazon Textract processes a PDF, it extracts all text, including elements like headers, footers, and page numbers. This extra information can be “noisy” and reduce the effectiveness of the output.

The process of converting a document’s 2D layout into a clean, one-dimensional string of text is called linearization.

The AmazonTextractPDFLoader gives you precise control over this process with the linearization_config parameter. You can use it to specify which elements to exclude from the final output.

The following example shows how to hide headers, footers, and figures, resulting in a much cleaner text block, for more advanced use cases see this AWS blog post.

from langchain_community.document_loaders import AmazonTextractPDFLoader
from textractor.data.text_linearization_config import TextLinearizationConfig

loader = AmazonTextractPDFLoader(
    "s3://amazon-textract-public-content/langchain/layout-parser-paper.pdf",
    linearization_config=TextLinearizationConfig(
        hide_header_layout=True,
        hide_footer_layout=True,
        hide_figure_layout=True,
    ),
)
documents = loader.load()

Using the AmazonTextractPDFLoader in a LangChain chain (e.g. OpenAI)

The AmazonTextractPDFLoader can be used in a chain the same way the other loaders are used. Textract itself does have a Query feature, which offers similar functionality to the QA chain in this example, which is worth checking out as well.

# You can store your OPENAI_API_KEY in a .env file as well
# import os
# from dotenv import load_dotenv

# load_dotenv()

# Or set the OpenAI key in the environment directly
import os

os.environ["OPENAI_API_KEY"] = "your-OpenAI-API-key"

from langchain.chains.question_answering import load_qa_chain
from langchain_openai import OpenAI

chain = load_qa_chain(llm=OpenAI(), chain_type="map_reduce")
query = ["Who are the authors?"]

chain.run(input_documents=documents, question=query)

' The authors are Zejiang Shen, Ruochen Zhang, Melissa Dell, Benjamin Charles Germain Lee, Jacob Carlson, Weining Li, Gardner, M., Grus, J., Neumann, M., Tafjord, O., Dasigi, P., Liu, N., Peters, M., Schmitz, M., Zettlemoyer, L., Lukasz Garncarek, Powalski, R., Stanislawek, T., Topolski, B., Halama, P., Gralinski, F., Graves, A., Fernández, S., Gomez, F., Schmidhuber, J., Harley, A.W., Ufkes, A., Derpanis, K.G., He, K., Gkioxari, G., Dollár, P., Girshick, R., He, K., Zhang, X., Ren, S., Sun, J., Kay, A., Lamiroy, B., Lopresti, D., Mears, J., Jakeway, E., Ferriter, M., Adams, C., Yarasavage, N., Thomas, D., Zwaard, K., Li, M., Cui, L., Huang,'