Our app uses Azure functions with Prompt Flow to call Azure OpenAI.
After a smooth deployment almost a month ago , we woke up to a message saying Production is down. The LLM calls don’t seem to be working.
A quick look at the app showed the Azure function was unloaded and the app showed 0 functions loaded. The logs showed a typical promptflow error
Exception: Promptflow may not installed correctly. If you are upgrading from ‘promptflow<1.8.0’ to ‘promptflow>=1.8.0’, please run ‘pip uninstall -y promptflow promptflow-core promptflow-devkit promptflow-azure’, then ‘pip install promptflow>=1.8.0’. Reach https://microsoft.github.io/promptflow/how-to-guides/faq.html#promptflow-1-8-0-upgrade-guide for more information.
We are on Promptflow version 1.17.2 so we were definitely not trying to upgrade.
Restarting, redeploying and recreating the function app didn’t work. Interestingly even creating a new instance of the function app in our stage environment gave the same error
So we had in our stage environment existing app X which was working, new App X1 with same code as X which gave the Promptflow error.
We narrowed it down to a Azure runtime version update that Microsoft rolled out over last few months and which finally made it to one of our app services ( and as per Murphy’s law,of course it won’t be rolled out in dev or stage first but directly in prod :-)) . This broke a dependency pipeline with Promptflow. We then got on a call with Microsoft support and confirmed our assumption!
So runtime version
currently has a dependency conflict with promptflow! Our other apps which are still working were at 4.1036.2.2.
Way forward
The support at Microsoft will be talking to the team and figuring out the next step on how to resolve this dependency ( maybe in their next update). Meanwhile we will pin down our so-far untouched app service plans to 4.1036.2.2 and deploy our code to one of the old app service plans ( all new ones are getting created with the new runtime version )
Promptflow is niche and quite new, not much help was available when we ran into this issue. Hope this helps a fellow programmer troubleshoot a similar issue 
Ah and this was a classic case of when Coding agents and tools couldn’t help up… good old logical thinking did 
. #AINeedsDevs !!
Update:- Here is the final fix ( with confirmation from Microsoft)
If deploying on Azure then apart from Promptflow also add the library called Promptflow[Azure].
A quick chat with chatgpt explains what promptflow-azure adds to the mix
🚀 What promptflow[azure] Adds on Top
When you install with pip install promptflow[azure], it includes Azure-specific extensions, enabling cloud-native development and MLOps.
Here’s what it unlocks:
✅ 1. Azure AI Studio Integration
-
Seamless flow deployment, testing, and management from Azure AI Studio.
-
View and manage runs in Azure UI.
✅ 2. Azure Machine Learning (Azure ML) Integration
-
Submit promptflows as Azure ML jobs.
-
Track runs, metrics, and outputs in Azure ML workspace.
-
Easily version and register flows in Azure.
✅ 3. Azure Dataset Support
-
Use Azure datasets (from Azure ML or storage) directly in your flows.
-
Manage datasets using Azure-native tools.
✅ 4. Authentication & Identity Integration
-
Automatically picks up Azure credentials (Azure CLI,
az login, managed identity, etc.). -
Secure access to resources (e.g., Azure OpenAI, storage, datasets).
✅ 5. CI/CD & MLOps Ready
-
Use
promptflow-azurein DevOps pipelines to deploy, test, and register promptflows in Azure. -
Integrate with GitHub Actions or Azure DevOps.
While none of this explains its relation with the new runtime we;ve learnt that it works to resolve this specific bug and maybe is more compatible with Azure deployments!
Hope this Helps!
Team Cennest