1/25/24 - Added support for language translation.
In this example, I provide an example of how leverage an Isolated Azure Function Blob Trigger that parses a Teams Meeting Transcript into a condensed format that is then used by AI to summarize the condensed transcript. I also expose an HTTP Trigger for Language Translation purposes. In the HTTP Post request you can pass a FileName, Container and target language and the API will translate the content into the target language. This example is leveraging the Semantic Kernel SDK so extending it's capibilities is super simple.
When calling the Translation API you need to pass the following JSON structure in the request body.
~~~
{
"fileName" : "Summary#1baa3e6e-46d7-4039-9581-871766800237#someemail@test.com#.txt",
"containerName" : "summary",
"targetLanguage" : "Spanish"
}
~~~
Important Note: I am not using the Azure AI Translation Service for the Language translation, I am simply using the GPT Language translation capabilities. You could leverage the AI Translation Service by using Semantic Kernel Plugins with a Native Function that calls the AI Translation services. Maybe I will implement that feature at a later date in this example.
The fileName is a valid file name that exists in the target container. The target contain needs to exist in the storage account associated with the BlobConnection environment variable.
The purpose for condensing the transcript is to reduce the number of AI tokens needed to summarize the meeting. Simply put, there are better formats that can be used when asking AI to summarize the meeting which makes it more efficient for AI to summarize and reduces token usage. If needed you could augment the TranscriptService class to chunk the data, especially if the file is large.
The API will open the file and read it's contents into a memory stream and pass it's contents to a Semantic Kernel Translation Plugin for translation and the translated text will be returned to the client.
- Isolated Azure Functions with a Blob Trigger
- Azure BlobServiceClient for writing to containers
- Semantic Kernel for AI Summarization purposes using Dependency Injection for the Kernel
- Semantic Kernel SKPrompts and Plugins
- StreamReader and Regex for file parsing
- Environment.GetEnvironmentVariable to read settings from local.settings.json and Azure Configuration Settings
-
You simply need to create 3 containers in a storage account; input, output and summary
-
Rename the local.settings.json.bak file to local.settings.json, then set the following values in local.settings.json file
"BlobConnection": "<Your connection string to the storage account>", "ApiDeploymentName": "<Your Deployment Name>", "ApiEndpoint": "<Your Endpoint>", "ApiKey": "<Your API Key"When deploying to Azure, you will need to make sure you create Configuration settings for BlobConnection, ApiDeploymentName, ApiEndPoint and ApiKey. When the funciton is running in Azure it readys the variables from Configuration settings and not the local.settings.json file.
-
The sample file found in the Data folder
-
The name of the input file is important as well and has the following format: transcript#1baa3e6e-46d7-4039-9581-871766800237#someemail@test.com#.txt.
**transcript** followed a # a **GUID** followed by a # then an email address, followed by a #, then the extension .TXT. -
The client that is pushing the file to an Azure Blob continer will need to generate a GUID and ensure the filename matches this format.
-
This format is using by the processing logic to extract the GUID and email address of the submitter and is used when writting new data to the output and summary containers
-
This format also allows the data to be written to various backing sources i.e. CosmosDB, SQL or other systems for history purposes
-
You can clone the repo and run it locally using Visual Studio.
-
Open the solution and press F5 to run it locally.
-
Navigate to the Azure Portal and drop/copy the transcript#1baa3e6e-46d7-4039-9581-871766800237#someemail@test.com#.txt file found in the Data folder to the Azure input container you created in the step above. If you set a breakpoint on the Blob Trigger, it will get hit once the file has been copied in the input container. You can step through the logic and watch the code condense the file and copy the condensed version to the output folder.
-
Now, navigate to the output container in the Azure portal and make note of the name of the file, it should be ReformattedTranscript#1baa3e6e-46d7-4039-9581-871766800237#someemail@test.com#.txt you will need it for the next step. If you navigate to the summary folder you will also find the AI summarized version of the transcript.
- Create a Canvas App that allows users to upload transcript to the input container using the file name formated outlined above. The Blob Trigger would be triggered and the summary would be written to the summary folder. The Canvas App could then retreive the summary from the summary folder.
- Create a Web App that uploads the transcript and reads the summary from the summary folder
- Basically, any client that can upload a file to Azure Storage can leverage this logic
-
Expose an HTTP Trigger that allows the client to pass a GUID that is used to retreive summaries from the summary container.
-
Add another Blob Trigger that allows for Document translation services i.e. convert documents written in English into other languages. In other words you can convert the meeting summary into other languages.
Below is an overview of the Architecture that could be leveraged with the Transcription Blob Trigger.
Example architecture that could be leveraged for the Language Translation HTTP Trigger.
