Inconsistent Adapter Results Between AWS Textract Console and Python SDK

Question

I'm experiencing a discrepancy when using AWS Textract with a custom adapter.

When I test my adapter through the AWS Console (Textract UI), it returns accurate results that align with the training. However, when I run the exact same query configuration using the Python SDK (boto3), I receive different and less accurate results, even though:
* The AdapterId and Adapter Version are exactly the same
* The document used is identical (same PDF file)
* The query text and aliases match exactly what was trained
* The SDK setup includes the correct AdaptersConfig and QueriesConfig

I’ve confirmed that:
* Version 2 of the adapter is active and working correctly in the Console
* Other versions or incorrect IDs are properly rejected by the API (showing validation works)

The same document and query setup produces different answers depending on whether I use the Console or the Python SDK.

```
pdf_bytes = download_file("doc-url-here")
pdf_path = save_temp_file(pdf_bytes, "pdf")

config = {
            "adapter_id": "a9c1efd06016",
            "queries": [
                {"Text": "chassi", "Alias": "chassi"},
                {"Text": "placa", "Alias": "placa"},
                {"Text": "data de vencimento", "Alias": "data de vencimento"},
            ]
        }

# Initialize Textract client
client = boto3.client("textract")

response = client.analyze_document(
            Document={'Bytes': pdf_bytes},
            FeatureTypes=['QUERIES'],
            QueriesConfig={'Queries': config["queries"]},
            AdaptersConfig={
                'Adapters': [
                    {
                        'AdapterId': config["adapter_id"],
                        'Version': '2'
                    }
                ]
            }
        )
```

M Zubair Bin Ramzan · Answer

1. Check Document Processing Parameters
Document Format: Ensure that the document being processed through the Python SDK is being uploaded in the exact same format and encoding as when you upload it through the Console. For example, ensure that the byte array (pdf_bytes) is correctly handled and matches the one used in the Console.

Image Quality: Double-check that the document's quality is consistent. If there are issues with OCR accuracy, these could be caused by minor discrepancies in the document format or quality between the Console and SDK.

2. Query Setup
In your config, the QueriesConfig is passed as {'Queries': config["queries"]}, but in the Console, queries are often pre-configured and tested with a specific set of conditions. Make sure that the queries are being processed identically both on the Console and in the SDK.

Query Aliases: Double-check that the query aliases (like "chassi", "placa", and "data de vencimento") are properly defined in both the Console and the SDK configuration.

3. AdaptersConfig and Version
Adapter Version: Ensure that the version ('Version': '2') is correctly referenced in the API call. Although you've confirmed that version 2 is active in the Console, verify if this version ID is properly associated with the specific query configurations you're using.

AdaptersConfig Structure: The AdaptersConfig structure in your request looks correct, but double-check if there's any subtle mismatch in the adapter configuration between the Console and the SDK request. Sometimes, the Console may implicitly set additional configurations that the SDK might not automatically pick up unless explicitly provided.

4. Consistency Between Console and SDK
Regional Differences: Verify that the API call in the SDK and the Console are both running in the same AWS region. Sometimes, different regions might have slight differences in the way features are processed or configured.

Feature Enablement: Double-check that the feature you are using (like QUERIES) is correctly enabled and available in both the Console and SDK.

Rate Limiting or Throttling: Ensure that you're not hitting any rate limits when using the SDK, which might cause discrepancies in results. Textract might perform differently under load.

5. Logging and Debugging
Enable debug logging in boto3 to capture the exact API request and response when calling analyze_document. This will give you more insight into how the request is being processed and may reveal any subtle differences in the API behavior.

import logging
boto3.set_stream_logger(name='botocore', level=logging.DEBUG)
This will log all the requests sent to AWS services and their responses, which can help you pinpoint discrepancies.

6. Alternative Approach: Test with CLI
You can test your exact configuration using the AWS CLI with the same AnalyzeDocument API call. This will help you confirm whether the issue is in the SDK or your specific Python implementation.

aws textract analyze-document \
  --document '{"Bytes": fileb://your-pdf-file.pdf}' \
  --feature-types "QUERIES" \
  --queries-config '{"Queries":[{"Text":"chassi","Alias":"chassi"},{"Text":"placa","Alias":"placa"}]}' \
  --adapters-config '{"Adapters":[{"AdapterId":"a9c1efd06016","Version":"2"}]}'
If the results from the CLI are the same as the Console, it might point to an issue in your SDK code or configuration.

7. Reaching Out to AWS Support
If none of these steps resolves the issue, it could be worth reaching out to AWS Support with the following details:

The exact differences in the results you are seeing.

The debug logs from the SDK.

Any other context about your configuration and Textract setup.

regards,
M Zubair
https://zeonedge.com

Inconsistent Adapter Results Between AWS Textract Console and Python SDK

Relevant content