Semantic Kernel Chronicles: Overview and Concepts
Semantic Kernel (SK) is an open-source software library that breaks down seemingly complex AI processes into more manageable components. If you are working on AI-related projects or plan to do so, this library is something you need to know about. This is because Semantic Kernel allows for faster and easier integration with AI by incorporating large language models (LLMs) like OpenAI and Hugging Face with currently supported programming languages such as C#, Python, and Java.
Developed by Microsoft, this tool makes the complex building blocks of AI more understandable and accessible. Semantic Kernel is an AI agent development kit. It is lightweight, open-source, and straightforward to use. You can integrate the latest AI models into your C#, Python, or Java codebase, enabling you to quickly manage complex AI processes and deliver enterprise solutions. It is also quite strong in terms of security, offering telemetry support, security filters, and the assurance of delivering responsible AI solutions.
Another significant advantage of Semantic Kernel is its forward-looking design. As AI models evolve rapidly, Semantic Kernel keeps pace with these changes. When new models are released, you can continue using the most up-to-date technology by making only the necessary changes, without needing to rewrite your entire codebase. This represents great convenience and cost savings for businesses. Additionally, by using your existing code as a plugin, you can flexibly integrate AI services and share these services with other developers.
To understand SK, which makes AI models more accessible and simplifies development processes, let’s discuss a few concepts highlighted by Microsoft:
Kernel: As the name suggests, it is the core, but to emphasize its difference from traditional local coding, the prefix “semantic” is used. We will build the entire structure on this core. The plugins, connectors, or configurations we want to use are registered on the kernel and made ready for use. The kernel coordinates tasks such as managing AI models, invoking functions, and processing results. Additionally, it provides integration with other modules and plugins. It also supports integrations like logging and telemetry, enabling performance and status checks of the application.
Agents: Agents are the fundamental building blocks of SK, AI-powered software components that interact with users and process requests. These agents receive instructions from users, perform the necessary operations, and report back the results. Within SK, agents can automate various functions by integrating with different modules and services.
Plugins: Plugins are a way to integrate existing code and functions into SK. They contain the functions necessary to perform specific tasks and work in harmony with other components of SK. This allows developers to add new AI features using existing APIs and code and utilize them within SK.
AI Services: AI Services enable SK to interact with various AI models and services. These services include various AI functions such as natural language processing (NLP), image recognition, and data analytics. SK uses these services to perform specific tasks and delivers them to users through agents.
Within Semantic Kernel, there are interfaces for the most popular AI tasks. In the table below, you can see the services that are supported by each of the SDKs.
Note: Some services are marked as experimental. This means that Microsoft has not yet completed the development of these services, is still making changes, and does not recommend them for production use. However, they are included in the SK infrastructure with C# support so that you can also experience them. It will be necessary to keep up with the updates :)
Planning: Planning relates to how SK organizes and manages tasks. It is used to coordinate multiple processes and AI models. This process involves determining the steps to follow, the services to use, and the expected outcomes. Planning ensures that AI solutions work effectively and efficiently.
Personas: Personas can be thought of as a type of “meta prompt” or “instruction” in Semantic Kernel. Personas are directives used to influence how an agent responds to stimuli. This allows you to influence how your agents plan tasks, generate responses, and interact with users. Microsoft defines these instructions as “personas” in Semantic Kernel because they closely resemble the concept of personas used in user experience design. Just as designers and UX researchers create personas to represent different types of users and the jobs they need to do, we can create personas to represent different types of agents and the tasks they are responsible for. By doing so, you can create agents that are more consistent, reliable, and predictable. This is especially important when building agents responsible for critical tasks or interacting with users in sensitive situations. With personas, your agents take on a specific role and behave accordingly, providing a more successful user experience.
I think it would be beneficial to share Microsoft’s best practices to keep in mind when creating Personas.
Additionally
To better understand the concepts of Agents and Plugins:
We mentioned that Agents are AI-powered components in Semantic Kernel that perform specific tasks and respond to user requests. These agents carry out the functions necessary to complete a given task. However, these functions are not always built directly into the agent; this is where Plugins come into play.
Plugins enable the integration of existing code and functions into SK. These plugins contain the functions needed to perform specific tasks, which can then be used by agents. For example, if an agent wants to fetch data from an API or automate a specific business process, it does so through a plugin.
Let’s examine a workflow example of an agent using a plugin:
• An Agent receives a command from a user. For example, the user wants to generate a report.
• To execute this command, the agent calls the necessary functions. If these functions do not already exist or if the agent wants to use a predefined function, a Plugin is utilized.
• The Agent calls the relevant Plugin to execute the required function. The Plugin performs the API call, data processing, or any other specific function and returns the results to the agent.
• The Agent then presents these results to the user or initiates the next step.
Through this connection, Agents can work flexibly and leverage existing functionality through Plugins to perform more complex tasks. This provides developers with the ability to reuse and extend their existing code while also empowering agents to become more robust and capable. By working together, Agents and Plugins enhance SK’s flexibility and strength, creating a synergy to automate complex business processes and more effectively utilize AI models.
