Context and Memory in SemanticKernel?

IT is a topic to write. Currently it isnt a draft. It is only some notes to work with.

Context

  • You can create on kernel like var context = kernel.CreateNewContext();
  • Then you can add values to it: like var history = ""; context.Variables["history"] = history;
  • Then between commends you can add it like:

history += $"\nUser: {input}\nMelody: {answer.GetValue()}\n"; context.Variables[“history”] = history; ```

  • It will be added to prompt

Planner:

  • How it decides which plugin should be used?
  • Is the plan genereted by UI ?
  • I think It can be prepared manually or it can be created by UI
  • Within Semantic Kernel, you can invoke these functions either manually (see chaining functions) or automatically with a planner.
  • If you want to order them manually you can do sth like that:
 var output = await this._kernel.RunAsync(
        request,
        getNumbers,
        extractNumbersFromJson,
        MathFunction,
        createResponse
    );

  • SemanticFunction

  • is is a way to probmpt

    • it has a defined bechavior
    • with parameters in format {{VARIABLE_NAME}}
    • it has defined parameters like maxTokens, Temperatre, TopP

  • Within a plugin, you can create two types of functions: semantic functions and native functions.

  • At its core, a semantic function is just a prompt that is sent to an AI service along with any additional settings the AI service requires.

  • SemanticFunction is like predefined prompt with:

    • with:
      • skprompt.txt => prompt
        • It can be with parameters
      • config.json -> a configuration how to use it
        • but here are parameters too

  • Native - it is a part of code

    • They can be used to
      • save data
      • retrieve data
      • and perform any other operation that you can do in code
      • do math, because LLM is not good in math :)
      • add argument [SKFunction ]

Example:

  • Before everything:
    • Populate memory with your context data
  • Executing (a pipeline)
    • Get user data context with NativeFunction (do a request for example)
    • Add domaincontext with recall from with TextMemoryPlugin.
    • Then make it a bit shorter for example with a specific semantic function
  • Get user query
  • prapare a bit of prompt
  • create a pipeline

Memories

  • a way to broader context for your ask
  • we can fed it. There are three the most populaor ways:
  • Conventional key-value pairs
  • Conventional local-storage
  • Semantic memory search
  • Q: Should I use it only for the user history or for the whole context?
    • Q: Is there a different place to store data?
  • It is jus an abstraction over the persisttation layer

SemanticFunction is important part of SematicKernel

Plugin:

comments powered by Disqus

Are you still here? Subscribe for more content!