Unknown error in sending http request with Azure Cognitive Services

fabsese 5

On my testing machine (Github codespace) everything works, but when I publish my changes and update my main project (Server located in Brazil), I get the following:

Error reason: ImageAnalysisErrorReason.CONNECTION_FAILURE
Error code: -1
Error message: Unknown error in sending http request

Additionally, if I make a Postman request to the same endpoint or I navigate to the endpoint in my browser, I get:

{
    "error": {
        "code": "404",
        "message": "Resource not found"
    }
}

I would expect the response to be something along the lines of "Invalid authentication" instead of "Not found"

Some searching online led me to a stackoverflow issue suggesting my endpoint formatting was incorrect, my endpoint shown on the Azure dashboard is https://[resourceName].cognitiveservices.azure.com/

whereas my speech endpoint is

https://eastus.api.cognitive.microsoft.com/sts/v1.0/issuetoken

I'm wondering if its incorrect?? Each endpoint is formatted differently, and on the stackoverflow issue I notice the solution was to use an endpoint similar to my speech endpoint, but this has not worked either.

I am using Python with the azure-ai-vision sdk version 0.13.0b1

romungi-MSFT 45,476 Reputation points Microsoft Employee

2023-07-07T12:32:12.7333333+00:00

@fabsese Based on the error message mentioned, I would speculate that you are using your resource subscription key as header with one or both of your endpoints to authenticate your requests but it fails with connection or unknown http error whereas the same works locally or in your github codespace. This suggests that the header value or subscription key passed is stripped in your request from the client machine after publishing your code on a server.

Also, the endpoint with the browser is expected to throw the error as resource not found.

However, in your postman request if you add an header Ocp-Apim-Subscription-Key and its value as key the request should be successful if valid body is passed.

It might be easier to understand if you could add how, you are using the speech and vision resource. Since you want to use the token for speech request and subscription key with vision SDK.

Also, are you using a single multi service cognitive services resource instead?

fabsese 5

I only mentioned speech endpoint as its the most recent functional project that has a different access link than the current one, being that the region is specified with speech but not computer vision. Speech and computer vision are used in the same project but seperately with different keys.

Here is my code:

import azure.ai.vision as visionsdk
import os

service_options = visionsdk.VisionServiceOptions(endpoint='https://catbot-eyes.cognitiveservices.azure.com', key=os.getenv("VISION_API_KEY"))

vision_source = visionsdk.VisionSource(
    url=url)

analysis_options = visionsdk.ImageAnalysisOptions()

if text:
    analysis_options.features = (
        visionsdk.ImageAnalysisFeature.CAPTION |
        visionsdk.ImageAnalysisFeature.TEXT
    )
else:
    analysis_options.features = (
        visionsdk.ImageAnalysisFeature.CAPTION
    )

analysis_options.language = "en"
# analysis_options.gender_neutral_caption = True

image_analyzer = visionsdk.ImageAnalyzer(service_options, vision_source, analysis_options)

result = image_analyzer.analyze()

if result.reason == visionsdk.ImageAnalysisResultReason.ANALYZED:
    #Process response and output...

else:
    error_details = visionsdk.ImageAnalysisErrorDetails.from_result(result)
    message = "Analysis failed.\n\n"
    message += f"-Error reason: {error_details.reason}\n"
    message += f"-Error code: {error_details.error_code}\n"
    message += f"-Error message: {error_details.message}\n"
    print(message)

I also included the full endpoint, figured that since it uses authentication it should be safe to post

romungi-MSFT 45,476 Reputation points Microsoft Employee

2023-07-11T06:34:31.9866667+00:00

@fabsese You are using the environment variable to get the key. Are you setting the key on your server too? Try adding the key directly for testing purposes when you publish to the server and check if the request is successful. I suspect the variable is not getting set when you are running your code on the server.
fabsese 5 Reputation points

2023-07-11T20:04:04.54+00:00

@romungi-MSFT The env is being loaded from a .env file using python-dotenv then os.

Valid key directly in the file on Codespaces is successful, valid key directly in the file on server and it fails 404.

Ports and firewall are open too.

fabsese 5

If I make an http request using the requests library, it works perfectly only when adding /vision/v3.0/analyze to the endpoint url. However, if I add /vision/v3.0/analyze to the SDK url, the entire process freezes indefinetly at the result = image_analyzer.analyze() line.

Is this a bug in the Azure SDK that I encountered? Plus an invalid/incomplete access URL being displayed on the Azure dashboard..

Successful code for reference:

import azure.ai.vision as visionsdk

service_options = visionsdk.VisionServiceOptions(endpoint='https://catbot-eyes.cognitiveservices.azure.com/vision/v3.0/analyze',
                                        key='removed')

vision_source = visionsdk.VisionSource(
    url='https://media.istockphoto.com/id/1093110112/photo/picturesque-morning-in-plitvice-national-park-colorful-spring-scene-of-green-forest-with-pure.jpg?s=612x612&w=0&k=20&c=lpQ1sQI49bYbTp9WQ_EfVltAqSP1DXg0Ia7APTjjxz4=')

analysis_options = visionsdk.ImageAnalysisOptions()

analysis_options.features = (
    visionsdk.ImageAnalysisFeature.CAPTION
)

analysis_options.language = "en"
image_analyzer = visionsdk.ImageAnalyzer(service_options, vision_source, analysis_options)

result = image_analyzer.analyze()

if result.reason == visionsdk.ImageAnalysisResultReason.ANALYZED:

    message = f"{result.caption.content.capitalize()}\nConfidence {result.caption.confidence*100:.4f}%\n\n"
    print(message)

else:

    error_details = visionsdk.ImageAnalysisErrorDetails.from_result(result)
    message = "Analysis failed.\n\n"
    message += f"-Error reason: {error_details.reason}\n"
    message += f"-Error code: {error_details.error_code}\n"
    message += f"-Error message: {error_details.message}\n"
    print(message)

romungi-MSFT 45,476 Reputation points Microsoft Employee

2023-07-12T05:04:18+00:00

@fabsese With the SDK the API path /vision/v3.0/analyze is not required as the SDK should handle this while making the call. The endpoint would be the same as you see on Azure portal from keys + Endpoints blade. Sometimes a reinstall of the SDK helps to fix any issues with the client. You could try that and report any issues on SDK repo if you still see this failing. Thanks!!
Sunil Sattiraju 6 Reputation points Microsoft Employee

2023-07-14T01:35:45.58+00:00

having the same problem, is there a solution for this?
Jonas Heschl 20 Reputation points

2023-08-14T14:38:59.1566667+00:00

I have the same issue working with Computer Vision using C# .NET Core client libraries. I have opened a Stackoverflow thread accordingly. Perhaps this can help someone here. https://stackoverflow.com/questions/76899163/azure-computer-vision-unknown-error-in-sending-http-request-when-running-loc
fabsese 5 Reputation points

2023-08-14T16:20:12.96+00:00

Terribly sorry I forgot entirely about this thread, it seems its an issue with the SDK, my solution was to use an HTTP request instead with the following URL:
https://[endpoint].cognitiveservices.azure.com/vision/v3.0/analyze
and your key as a header:
"Ocp-Apim-Subscription-Key": subscription_key

If you use a different cognitive service, you will have to look for its respective /vision/v3.0/analyze equivalent.

Edit: I do think the SDK should have some fallback protection for when an invalid endpoint URL is provided, for example in my previous comment where setting the URL to have /vision/v3.0/analyze at the end, it freezes indefinitely with no (reasonable) timeout or explanation.

Share via

Unknown error in sending http request with Azure Cognitive Services

Your answer