[Python] Import semantic skills

[lightningRalf]

Hello,
I am no programmer, but I want to understand the new technology, so please forgive me if these questions are out of place.

My understanding:

At the moment we got the ImportSkills(ExtendsKernel) class. In there is a function to import from a directory.

Question:

Are there other functions planned? Like calling from normal databases like SQLite, or even vectordatabases like elastic search, or similar?

Why do I ask?

Because I think with prompt engineering there could done a lot, and all these tiny changes require a new skill. So we end up with 10s of 100s of skills just for prompting.
And I think there will be demand for analysing the behavior of these prompts in different situations.

Usecase:

Imaging the following usecase:
Analysing different prompts to outcomes while using skills and visualization, like Aim.
And feeding back the outcomes automatically back into the database / vectorstore.

Would greatly appreciate some feedback.
Greetings from Germany

[timlaverty]

Hi lightningRalf, Our mental model is folks will add the skills they need. If you're writing a solution and you need SQLite (or other store access) you could add a skill to do what you need. You don't need to modify the kernel - as you're describing, it's straightforward to add a skill and load it into the kernel at runtime.

For your usecase: I'm not super familiar w/ Aim but I think it'd be reasonable to pull inputs, outputs of LLMs and tools, leverage SK to generate embeddings, and push those into a store.

[lightningRalf]

Thanks for getting back to me.

I understood the following: set up a second SK that pulls the inputs and outputs everytime these are initiated and then generate embeddings.

I want the skills library to fill it self, with a "self-learning algorithm".
Something along this line: [vague try to paint what I mean]

So for example I tell agent1: try a few-shots prompt constiting of skills in this group, go (for example) for five whole rounds.
Than I should set up another SK agent2 that compares the prompts and the results, Weighs them and puts these informations in a memory base.

But with these settings agent1 is not able to learn during his process from the first iterations.

Hope this thought process is somehow followable?

[dluc]

possibly relevant: #506