Azure Document Intelligence - Mismatch Between Trained API Version and Response API Version

Question

Hello,

I have trained a Custom Neural Model under Document Intelligence API Version 2023-07-31.
When performing inference via the API, the "apiVersion" field is not consistent with the trained API version. Inference returns a response with API version 2024-02-29-preview. The problem is consistent across custom neural models and prebuilt-layout. The issue persists when specifying a model version directly in the HTTP request.

Can anyone recreate this behavior, and could MS please advise on or fix this issue?

Test Case 1: Custom Neural Model

Model:

Code:

import os
import yaml
from azure.core.credentials import AzureKeyCredential
from azure.ai.documentintelligence import DocumentIntelligenceClient

def load_azure_config(directory="."):
    filename = "azure_config.yaml"
    filepath = os.path.join(directory, filename)
    with open(filepath, "r") as stream:
        try:
            azure_config = yaml.safe_load(stream)
        except yaml.YAMLError as exc:   
            print(exc)
    return azure_config

def get_doc_client(AzureKeys):
    key = AzureKeys["DocumentIntelligence"]["KEY_1"]
    endpoint = AzureKeys["DocumentIntelligence"]["AZURE_ENDPOINT"]
    return DocumentIntelligenceClient(endpoint, AzureKeyCredential(key))

def analyze_document(filepath, doc_client, model_id):
    with open(filepath, "rb") as f:
        poller = doc_client.begin_analyze_document(
            model_id=model_id, 
            analyze_request=f, 
            content_type="application/octet-stream"
        )
    return poller.result()

# Main execution
if __name__ == "__main__":
    # Load Azure configuration
    AzureKeys = load_azure_config()['AzureKeys']

    # Set up the Document Intelligence client
    doc_client = get_doc_client(AzureKeys)

    # Define the input file and model ID
    input_file = r"C:\Users\xxxxxx\Downloads\AIAI2021 (1).pdf"
    model_id = "test-response"

    # Analyze the document
    result = analyze_document(input_file, doc_client, model_id)

    # Print the result (you can modify this to process the result as needed)
    print(result.as_dict())

Response:

{'apiVersion': '2024-02-29-preview', 'modelId': 'test-response', 'stringIndexType': 'textElements', 'content': 'Goldsmiths Research Online\nCitation\nGoldsmiths Research Online (GRO) is the institutional research repository for Goldsmiths, University of London\nOlaniyan, Rapheal; Stamate, Daniel and Pu, Ida. 2021. \'A Two-Step Optimised BERT-Based NLP Algorithm for Extracting Sentiment from Financial News\' ... }

Test Case 2: Prebuilt Layout Model

Model:

prebuilt-layout version 2024-07-31

Code:

import os
import yaml
import requests
import time

def load_azure_config(directory="."):
    """
    Loads Azure configuration from a YAML file.
    """
    filename = "azure_config.yaml"
    filepath = os.path.join(directory, filename)
    with open(filepath, "r") as stream:
        try:
            azure_config = yaml.safe_load(stream)
        except yaml.YAMLError as exc:
            print(exc)
    return azure_config


def analyze_document_rest(filepath, azure_keys, model_id, api_version="2023-07-31"):
    """
    Submits a document for analysis using the Azure Document Intelligence REST API and polls for results.

    Args:
        filepath (str): Path to the document file to analyze.
        azure_keys (dict): Azure keys and endpoint information.
        model_id (str): Document model ID (e.g., 'prebuilt-layout').
        api_version (str): API version to use.

    Returns:
        dict: Analysis results as a dictionary, or None if an error occurred.
    """
    endpoint = azure_keys["DocumentIntelligence"]["AZURE_ENDPOINT"]
    key = azure_keys["DocumentIntelligence"]["KEY_1"]

    # Correct the REST API URL
    url = f"{endpoint}/formrecognizer/documentModels/{model_id}:analyze?api-version={api_version}"

    headers = {
        "Ocp-Apim-Subscription-Key": key,
        "Content-Type": "application/octet-stream",
    }

    try:
        # Read the file and send the request
        with open(filepath, "rb") as file_data:
            response = requests.post(url, headers=headers, data=file_data)

        # Handle initial response
        if response.status_code == 404:
            print(
                "Error 404: Endpoint or resource not found. Check your endpoint, model ID, or API version.")
            return None
        elif response.status_code != 202:
            print(f"Error: {response.status_code}")
            try:
                print(response.json())  # Print error details if available
            except Exception:
                print(response.text)  # Fallback if JSON decoding fails
            return None

        # Poll for the result using the operation-location header
        operation_location = response.headers.get("Operation-Location")
        if not operation_location:
            print("Error: Operation-Location header not found.")
            return None

        print("Document analysis submitted. Polling for results...")

        # Polling for result
        while True:
            poll_response = requests.get(operation_location, headers={
                                         "Ocp-Apim-Subscription-Key": key})
            if poll_response.status_code != 200:
                print(f"Polling Error: {poll_response.status_code}")
                print(poll_response.text)
                return None

            poll_data = poll_response.json()
            status = poll_data.get("status")
            if status == "succeeded":
                print("Analysis succeeded.")
                return poll_data
            elif status == "failed":
                print("Analysis failed.")
                print(poll_data)
                return None

            # Wait and retry polling
            print("Waiting for 5 seconds before retrying...")
            time.sleep(5)

    except Exception as e:
        print(f"An error occurred: {e}")
        return None


# Load Azure configuration
azure_keys = load_azure_config()["AzureKeys"]

# Define the input file and model ID
input_file = r"C:\Users\xxxxxxx\Downloads\AIAI2021 (1).pdf"
model_id = "prebuilt-layout"

# Specify the API version
api_version = "2023-07-31"

# Analyze the document using REST API
result = analyze_document_rest(input_file, azure_keys, model_id, api_version)

Response:

{'apiVersion': '2024-02-29-preview',
 'modelId': 'prebuilt-layout',
 'stringIndexType': 'textElements',
 'content': 'Goldsmiths Research Online\nGoldsmiths '...}

Answer 1

@Jack Alvarez Welcome to Microsoft Q&A Forum, Thank you for posting your query here!

.

This is expected behavior.

.

The API version (2024-02-29-preview )and model version ( 2023-07-31)are different concepts. A new API version indicates changes to the API since the previous version. This could involve the addition of new APIs or updates to specific features, such as model improvements (model version changes). However, it doesn't imply that all models have been updated.

.

Hope this helps. If you have any follow-up questions, please let me know. I would be happy to help.