The Unstructured Python SDK client allows you to send one file at a time for processing by the Unstructured Partition Endpoint. To use the Python SDK, you’ll first need to set an environment variable named UNSTRUCTURED_API_KEY, representing your Unstructured API key. To get your API key, do the following:
  1. If you do not already have an Unstructured account, sign up for free. After you sign up, you are automatically signed in to your new Unstructured Starter account, at https://platform.unstructured.io.
    To sign up for a Team or Enterprise account instead, contact Unstructured Sales, or learn more.
  2. If you have an Unstructured Starter or Team account and are not already signed in, sign in to your account at https://platform.unstructured.io.
    For an Enterprise account, see your Unstructured account administrator for instructions, or email Unstructured Support at support@unstructured.io.
  3. Get your Unstructured API key:
     a. After you sign in to your Unstructured Starter account, click API Keys on the sidebar.
    For a Team or Enterprise account, before you click API Keys, make sure you have selected the organizational workspace you want to create an API key for. Each API key works with one and only one organizational workspace. Learn more.
    b. Click Generate API Key.
    c. Follow the on-screen instructions to finish generating the key.
    d. Click the Copy icon next to your new key to add the key to your system’s clipboard. If you lose this key, simply return and click the Copy icon again.

​
Installation

Before using the SDK to interact with Unstructured, install the library:
Python
pip install unstructured-client
The SDK uses semantic versioning and major bumps could bring breaking changes. It is advised to pin your installed version. See the migration guide, later on this page, for breaking change announcements.

​
Basics

Let’s start with a simple example in which you send a PDF document to the Unstructured Partition Endpoint to be partitioned by Unstructured.
Python
import os, json

import unstructured_client
from unstructured_client.models import operations, shared

client = unstructured_client.UnstructuredClient(
    api_key_auth=os.getenv("UNSTRUCTURED_API_KEY")
)

filename = "PATH_TO_INPUT_FILE"

req = operations.PartitionRequest(
    partition_parameters=shared.PartitionParameters(
        files=shared.Files(
            content=open(filename, "rb"),
            file_name=filename,
        ),
        strategy=shared.Strategy.VLM,
        vlm_model="gpt-4o",
        vlm_model_provider="openai",
        languages=['eng'],
        split_pdf_page=True,            # If True, splits the PDF file into smaller chunks of pages.
        split_pdf_allow_failed=True,    # If True, the partitioning continues even if some pages fail.
        split_pdf_concurrency_level=15  # Set the number of concurrent request to the maximum value: 15.
    ),
)

try:
    res = client.general.partition(
        request=req
    )
    element_dicts = [element for element in res.elements]
    
    # Print the processed data's first element only.
    print(element_dicts[0])

    # Write the processed data to a local file.
    json_elements = json.dumps(element_dicts, indent=2)

    with open("PATH_TO_OUTPUT_FILE", "w") as file:
        file.write(json_elements)
except Exception as e:
    print(e)

​
Async partitioning

The Python SDK also has a partition_async. This call is equivalent to partition except that it can be used in a non blocking context. For instance, asyncio.gather can be used to concurrently process multiple files inside of a directory hierarchy, as demonstrated here: 
import asyncio
import os
import json
import unstructured_client
from unstructured_client.models import shared

client = unstructured_client.UnstructuredClient(
    api_key_auth=os.getenv("UNSTRUCTURED_API_KEY")
)

async def call_api(filename, input_dir, output_dir):
    req = {
        "partition_parameters": {
            "files": {
                "content": open(filename, "rb"),
                "file_name": os.path.basename(filename),
            },
            "strategy": shared.Strategy.VLM,
            "vlm_model": "gpt-4o",
            "vlm_model_provider": "openai",
            "languages": ['eng'],
            "split_pdf_page", True,            # If True, splits the PDF file into smaller chunks of pages.
            "split_pdf_allow_failed": True,    # If True, the partitioning continues even if some pages fail.
            "split_pdf_concurrency_level": 15  # Set the number of concurrent request to the maximum value: 15.
        }
    }

    try:
        res = await client.general.partition_async(
            request=req
        )
        element_dicts = [element for element in res.elements]
        json_elements = json.dumps(element_dicts, indent=2)

        # Create the output directory structure.
        relative_path = os.path.relpath(os.path.dirname(filename), input_dir)
        output_subdir = os.path.join(output_dir, relative_path)
        os.makedirs(output_subdir, exist_ok=True)

        # Write the output file.
        output_filename = os.path.join(output_subdir, os.path.basename(filename) + ".json")
        with open(output_filename, "w") as file:
            file.write(json_elements)

    except Exception as e:
        print(f"Error processing {filename}: {e}")

async def process_files(input_directory, output_directory):
    tasks = []

    for root, _, files in os.walk(input_directory):
        for file in files:
            if not file.endswith('.json'):
                full_path = os.path.join(root, file)
                tasks.append(call_api(full_path, input_directory, output_directory))

    await asyncio.gather(*tasks)

if __name__ == "__main__":
    asyncio.run(process_files(
        input_directory=os.getenv("LOCAL_FILE_INPUT_DIR"), 
        output_directory=os.getenv("LOCAL_FILE_OUTPUT_DIR")
    ))

​
Page splitting

In order to speed up processing of large PDF files, the split_pdf_page* parameter is True by default. This causes the PDF to be split into small batches of pages before sending requests to the API. The client awaits all parallel requests and combines the responses into a single response object. This is specific to PDF files and other filetypes are ignored. The number of parallel requests is controlled by split_pdf_concurrency_level*. The default is 8 and the max is set to 15 to avoid high resource usage and costs. If at least one request is successful, the responses are combined into a single response object. An error is returned only if all requests failed or there was an error during splitting.
This feature may lead to unexpected results when chunking because the server does not see the entire document context at once. If you’d like to chunk across the whole document and still get the speedup from parallel processing, you can:
  • Partition the PDF with split_pdf_page set to True, without any chunking parameters.
  • Store the returned elements in results.json.
  • Partition this JSON file with the desired chunking parameters.
Python
req = operations.PartitionRequest(
    partition_parameters=shared.PartitionParameters(
        files=shared.Files(
            content=file.read(),
            file_name=filename,
        ),
        strategy=shared.Strategy.VLM,
        vlm_model="gpt-4o",
        vlm_model_provider="openai",
        split_pdf_page=True,            # If True, splits the PDF file into smaller chunks of pages.
        split_pdf_allow_failed=True,    # If True, the partitioning continues even if some pages fail.
        split_pdf_concurrency_level=15  # Set the number of concurrent request to the maximum value: 15.
    )
)
res = client.general.partition(
    request=req
)

​
Customizing the client

​
Retries

You can also change the defaults for retries through the retry_config* parameter when initializing the client. If a request to the API fails, the client will retry the request with an exponential backoff strategy up to a maximum interval of one minute. The function keeps retrying until the total elapsed time exceeds max_elapsed_time*, which defaults to one hour. However, you can override these defaults, for example as follows:
Python
# ...
from unstructured_client.utils import RetryConfig, BackoffStrategy

client = UnstructuredClient(
    api_key_auth=os.getenv("UNSTRUCTURED_API_KEY")
    retry_config=RetryConfig(
        strategy="backoff",
        retry_connection_errors=True,
        backoff=BackoffStrategy(
            initial_interval=500,     # 0.5 seconds: 500 milliseconds
            max_interval=60000,       # 1 minute: 60 seconds * 1000 milliseconds
            exponent=1.5,
            max_elapsed_time=900000,  # 15 minutes: 15 * 60 seconds * 1000 milliseconds
        ),
    )
)
Available RetryConfig settings include:
SettingDescription
strategyThe strategy to use for retries. The only supported value is backoff.
retry_connection_errorsTrue to retry on connection errors.
BackoffStrategy.initial_intervalAfter the first error, wait the specified number of milliseconds before retrying.
BackoffStrategy.max_intervalThe maximum wait time, specified in milliseconds between retries.
BackoffStrategy.exponentAfter each retry, increase the wait time exponentially by the specified factor.
BackoffStrategy.max_elapsed_timeStop retrying after the specified number of milliseconds.

​
Disabling SSL validation

If you disable SSL validation, requests will accept any TLS certificate presented by the server and ignore hostname mismatches and/or expired certificates, which will make your application vulnerable to man-in-the-middle (MitM) attacks. Only set this to False for testing.
Python
http_client = requests.Session()
http_client.verify = False

client = UnstructuredClient(
    client=http_client,
    ...
)

​
Handling the response

The partition response defaults to a dict format that can be converted to Unstructured elements with the elements_from_dicts utility function as seen below. Otherwise, the API response can be sent directly to your vector store or another destination.
Python
from unstructured.staging.base import elements_from_dicts

# ...

if res.elements is not None:
    elements = elements_from_dicts(response.elements)

​
Parameters & examples

The parameter names used in this document are for the Python SDK, which follow snake_case convention. The JavaScript/TypeScript SDK follows camelCase convention. Other than this difference in naming convention, the names used in the SDKs are the same across all methods.
  • Refer to the API parameters page for the full list of available parameters.
  • Refer to the Examples page for some inspiration on using the parameters.

​
Migration guide

There are breaking changes beginning with Python SDK version 0.26.0. If you encounter any errors when upgrading, please find the solution below. If you see the error: AttributeError: 'PartitionParameters' object has no attribute 'partition_parameters' Before 0.26.0, the SDK accepted a PartitionParameters object as input to the sdk.general.partition function. Beginning with 0.26.0, this object must be wrapped in a PartitionRequest object. The old behavior was deprecated in 0.23.0 and removed in 0.26.0. 
# Instead of:
from unstructured_client.models import shared

req = shared.PartitionParameters(
    files=files,
)

resp = s.general.partition(
    request=req
)

# Switch to:
from unstructured_client.models import shared, operations

req = operations.PartitionRequest(
    partition_parameters=shared.PartitionParameters(
        files=files,
    )
)

resp = s.general.partition(
    request=req
)
If you see the error: TypeError: BaseModel.__init__() takes 1 positional argument but 2 were given Beginning with 0.26.0, the PartitionRequest constructor no longer allows for positional arguments. You must specify partition_parameters by name. 
# Instead of:
req = operations.PartitionRequest(
    shared.PartitionParameters(
        files=files,
    )
)

# Switch to:
req = operations.PartitionRequest(
    partition_parameters=shared.PartitionParameters(
        files=files,
    )
)
If you see the error: TypeError: General.partition() takes 1 positional argument but 2 were given Beginning with 0.26.0, the partition function no longer allows for positional arguments. You must specify request by name. 
# Instead of:
resp = s.general.partition(req)

# Switch to:
resp = s.general.partition(
    request=req
)