Adventures in Semantic Kernel

Extensive Interactive Demo of the Semantic Kernel SDK

Welcome to Adventures in Semantic Kernel, your interactive guide to exploring the functionalities of Microsoft's AI Orchestration library, Semantic Kernel. Dive into hands-on experiences ranging from dynamic plan generation and Agent building for a dynamic chat experience to memory management and tokenization. This isn't just a passive learning experience; you'll get to actively experiment with these features to understand their cohesive interactions. Try it out here

About Semantic Kernel

Originally developed by Microsoft, Semantic Kernel aims to democratize AI integration for developers. While the project benefits from open-source contributions, its core mission is to simplify the integration of AI services with app code. It comes equipped with a smart set of connectors that essentially act as your app's "virtual brain", capable of executing LLM prompts, native code or external REST Apis.

Configurations

Current configuration will work for all the main features of the demo, and for most (though not all) plugins. However, several KernelSyntaxExamples will require config values for specific resources (e.g. Pinecone, Chroma, Weaviate, etc.) not available by default. Any service config highlighted in red is missing values that will need to be added for the associated sample to work.

You don't need to supply an OpenAI api key for most of the demo features, but if you want to use a gpt-4 model (or if you want to change the default service to Azure OAI), you will need to supply an api key in the OpenAIConfig or AzureOpenAIConfig section. Note: All configurations added/changed are encrypted and saved to your browser's local storage so they can be loaded across sessions while remaining secure.

Application features

Samples

View, modify, and execute dotnet examples. Examples are from KernelSyntaxExamples with small modifications.

Execute Function

Select a single plugin from a large variety of native, prompt and external plugins, then execute a function from that plugin.

Build Agent

Build a simple agent by providing a persona and collection of plugins used together with OpenAI Function Calling.

Build Agent Group Chat

Build a group chat comprised of ChatCompletionAgents using AgentGroupChat

Build Planner

Select plugins and functions to build and execute your own:

• OpenAI Function Calling Agent
• Handlebars planner (Still available, but usage is discouraged in favor of Function Calling. See Planners vs Function Calling for more information.)
• Stepwise planner (Still available, but usage is discouraged in favor of Function Calling. See Planners vs Function Calling for more information.)

Custom Examples

Web Chat Agent

Chat with the web using Bing search and a scrape-and-summarize plugin

Wikipedia Chat Agent

Chat with the Wikipedia articles using Wikipedia Rest API

C# REPL Agent

Use natural language prompts to generate and execute c# code

• Generate and execute a c# console application using prompts.
• Generate and execute c# line-by-line using Roslyn c# scripting api.

Dnd Story Agent

Example of a Stepwise Planner at work. Planner has access to the D&D5e Api plugin and multiple prompt plugins. It uses these to create and execute a plan to generate a short story.

• Leverages a native plugin from a Razor Class Library AskUserPlugin to provide user interaction during plan execution

SK Memory

Vector Playground

Play around with embeddings and similarities using your own or generated text snippets

SK + Custom Hdbscan Clustering

See how embeddings can be used to cluster text items, and then generate a title and summmary for each cluster using prompt plugins

Tokens

Chunking and Tokenization

• Generate or add text, set the text chunking parameters, and then see the Semantic Kernel TextChunker work
• Search over chunked text to see how the TextChunker can be used to improve search results

Tinker with Tokens

See how input text translates into tokens. Select specific tokens to set the LogitBias for a chat completion request/response.