This week I’ve been researching Text-to-SQL (also known as Natural Language to SQL), below are my study notes to compile all the resources I’ve found on the topic to date. There is also a corresponding blog entry that walks through a code example.

NOTE: I am approaching this topic specifically looking at how it can be used to extend usage scenarios in a RAG application.

Background

Text-to-SQL (or Natural Language to SQL) is a pattern where the objective is to have an LLM generate SQL statements for a database using natural language.

Last night was the Season of AI presentation. We started with Bill Wilder presenting the fundamentals of Generative AI and quick introduction to Azure AI Studio, then I finished up with a .NET code walkthrough implement Retrieval Augmented Generation (RAG) using Semantic Kernel.

It was nice to see a lot of regular faces and meet several new people.

Demo Code

The demo code is on my GitHub repo BostonAzure-June2024 under a subdirectory.

This is the fifth C# demo in The RAG Demo Chronicles (Blog Series) and has the most extensive use of Semantic Kernel out of all the demos I’ve reviewed.

The use of Retrieval Augmented Generation (RAG) is different with this project than the other demos I’ve reviewed - mainly because RAG is just one of its features.

With this demo, I also took the time to configure the optional authentication so I could play with the MS Graph plugin … and WOW! Just adding the web searcher and MS Graph plugin really make this a powerful demo app. It gives me a lot of ideas for my own projects.

UPDATE: The Bing Search APIS are being deprecated. See the announcement for more details: Bing Search APIs retiring on August 11, 2025. I have updated the post and corresponding Github code to use the Brave connector instead.

A couple of weeks ago I thought I’d written my last of these blogs, mainly due to me getting more in depth with Semantic Kernel. However, after I watched Will Velida’s video Using Bing Search API in the Semantic Kernel SDK … I couldn’t help but wonder what the API calls were behind the scenes. Will does a great job at explaining how to use the plugin and the Bing resource needed to make calls to the search API, so I won’t get into that part of it - I want to focus on the usefulness and API calls made by the plugin.

UPDATE: Planners have been deprecated. See this Semantic Kernel blog for more detail: Semantic Kernel: Package previews, Graduations & Deprecations

Last week in the Semantic Kernel Hello World Planners Part 1 entry, I used the Handlebars planner to implement the sample Hello World functionality and then looked at the token difference between using a saved plan vs. generating a plan. In this entry I use the Function Calling Stepwise Planner to create the sample Hello World functionality and compare it to the implementation in the Semantic Kernel Hello World Plugins Part 3 entry.

UPDATE: Planners have been deprecated. See this Semantic Kernel blog for more detail: Semantic Kernel: Package previews, Graduations & Deprecations

A few weeks ago in the Semantic Kernel Hello World Plugins Part 3 blog entry, I showed how to use OpenAI Function Calling. The last half of that entry was all about how to view the response and request JSON going back and forth to OpenAI, which detailed four API calls. In this entry I look at using the Handlebars Planner to accomplish the same functionality. Then I’ll show the request and response JSON for both using a saved plan as well as having the LLM create a plan and end with a token usage comparison.

Last Thursday night I spoke at the Memphis Azure User Group, it was nice to meet some people in person and see how excited others are about finding valuable ways to work GenAI into their applications.

I also gave my slide deck a completely new look (Memphis themed via Bing/create):

Talk: Getting Started with Retrieval Augmented Generation (RAG)

The presentation pdf can be downloaded here.

Since the presentation was hybrid, there were not as many questions as the other two times I’ve given the talk … or that is my guess at the reason why it was so quiet. That also means I should have put that third demo back in the presentation (I took a demo out after the first version of the talk since there were so many questions and needed to free up time for Q&A). Next time I’ll have it ready to use in case I need it.

Last week I blogged Part 2 showing the creation of a native function plugin, in this post I want to take that native function a step further and use the OpenAI Function calling. This will allow us to not provide the current date when making the call to get a historic daily fact and have OpenAI call a function to get the current date.

I’ve added the HelloWorld.Plugin3.Console project to the GitHub repo for the code in this blog entry.

Two weeks ago I blogged Part 1, in which I moved the prompt to a prompt template. In this part, I implement a native function that will take in the current date and make the call to the LLM.

I’ve put the code for this blog in the HelloWorld.Plugin2.Console project in the same repo as the other SK entries: semantic-kernel-getting-started.

Semantic Kernel Plugin: Native Function

There is a good Microsoft Learn module: Give your AI agent skills that walks you through the details of what a native function is and how to implement them.

This past weekend was Boston Azure’s Edition of the annual Global Azure Bootcamp. This year we focused on AI and hands-on-labs.

The odd thing about when we scheduled the meetup was we had a lot of people sign up for the group just to rsvp - before most of the existing members had gotten around to rsvp’ing. We did not expect that. It is a mystery as how they heard about the event so quick. We only had one room reserved, so there really was a hard cap on how many people we could let in.