Add the PromptFlow from "On Your Data" and check it works with our index

[gaurarpit]

Required by #406

Description

Create and deploy a flow which mimics the current converstation flow.

Tasks

• Spike replicating the behavior in prompt flow and how to deploy
• Create flow
• Deploy the flow via Bicep - feat: Add prompt flow deployment #1003

[adamdougal]

Update 20th May:

• Currently working on spiking the implementation
• We've successfully replicated orchestration flow in prompt flow
• We've successfully exported the flow and run this locally using pf flow serve ...
• We're hitting an issue when trying to build the flow using pf flow build...
  • We need this to be able to deploy it to Azure App Service
  • The issue is that the Azure AI Search connection is not being picked up by the pf flow build command
  • We've raised an issue here as I believe it is a bug with the pf tool [BUG] AI Search connection in mlindex_content not detected microsoft/promptflow#3313
  • We'll keep digging for solutions

[adamdougal]

Further Update 20th May:

• I've managed to get around the issue with the AI Search connection not being picked up by adding connection: aisearch_connection to the lookup node
• I'm now running into another issue where running the built executable works, but sending queries through it results in a 401 -Unauthorized. Access token is missing, invalid, audience is incorrect (https://cognitiveservices.azure.com), or have expired. error when calling OpenAI.
  • I've checked the Key is correct and all is good
  • I'll continue to investigate

[adamdougal]

Update 21st May:

• After receiving guidance from the prompt flow team, I have been able to deploy the flow to an AML endpoint
• Discussed this with the team and this is more desirable than deploying to an App Service as it's closer to where the prompt flow is defined and experimentation is run
• Next step is to work out how to deploy an AML endpoint using bicep

[adamdougal]

Update 24th May:

• I've got the workspace and connections deploying successfully using Bicep
• However, it does not seem to be possible to deploy a flow and a corresponding endpoint using Bicep
• I've explored using deployment scripts, but these run on a container in Azure, it's not trivial to install the required dependencies (pfazure) and it does not have access to local files.
• As a fallback, we could look at using a azd postprovision hook, which runs locally to deploy the flow. The main downside of this, is that there would be a manual step for users who deploy using one click deploy.
• I'll discuss with the team to decide next steps.