kernel-memory in relationship with semantic-kernal

[liammoat]

I would like to better understand the relationship between kernel-memory and semantic-kernal.

I understand this project targets the specific requirement of "talk to my data" and, from my exploration, provides greater functionality compared to SK's Memory skills. It also works great as a standalone project, with capabilities to index, query, and provide natural language summarisation etc.

The README says:

Designed for seamless integration as a Plugin with Semantic Kernel

However, I cannot see how this is intended to work from an implementation perspective. All the samples use kernel-memory directly, without the need for SK. Are there any samples for this scenario?

[lanyeeee]

For the sake of understanding
let's say you're using Qdrant as a backend
import a pdf about the use of nginx
and then ask what nginx is, it works like this:

When you import file:
pdf file after tokenize, will be split into multiple partition
each partition will be sent to the LLM for embedding
and then each partition data with its corresponding embedding data will be saved into the Qdrant

When you ask:
your question will be sent to LLM for embedding
the embedding data of the question will be used to query in Qdrant to find out the partitions related to the question
and then the partitions will be written as facts in the prompt
like this

Facts:
{{$facts}}
======
Given only the facts above, provide a comprehensive/detailed answer.
You don't know where the knowledge comes from, just answer.
If you don't have sufficient information, reply with 'INFO NOT FOUND'.
Question: {{$input}}
Answer: 

and finally the prompt will be sent to LLM to get the answer.

The bold parts are done by kernel-memory, other parts are done by Semantic Kernel

[crickman]

Thank you for this question.

Chat-Copilot (https://github.com/microsoft/chat-copilot) is an example of an aplication that utilizes kernel-memory in conjunction with semantic-kernel.

For advanced application case, you may have the need to talk your data along with other prompt, planner, plug-in, or agent based activities.

[dluc]

hi @liammoat I suggest looking at Kernel Memory as a Plugin that can be used with Semantic Kernel and SK planners. SK includes a very basic Memory plugin that needs to be maintained in 3 languages and misses some important features like importing files, automatic chunking, RAG, asynchronous data ingestion, etc.

The plugin for C# is almost ready, see here, there's also one example: #146

// ...

var memory = new MemoryWebClient("http://127.0.0.1:9001/");

await memory.ImportDocumentAsync("mydocs-NASA-news.pdf");

kernel.ImportFunctions(new MemoryPlugin(memory), "memory");

var skPrompt = """
Question to Kernel Memory: {{$input}}

Kernel Memory Answer: {{memory.ask $input}}

If the answer is empty say 'I don't know' otherwise reply with a preview of the answer, truncated to 15 words.
""";

var oracle = kernel.CreateSemanticFunction(skPrompt);

KernelResult answer = await kernel.RunAsync("any news about Orion?", oracle);

// Answer: Yes, NASA has invited media to see the new test version of the Orion spacecraft...

answer = await kernel.RunAsync("any news about Hubble telescope?", oracle);

// Answer: I don't know.

[liammoat]

Thanks @dluc - that is exactly what I was looking for. 😊