Enhancement: a personal context management area for personalised AI experiences!

[danielrosehill]

Once I've got this idea jotted down I'll have sent over all my major feature ideas for the time being at least!

This is very similar to the datasets idea that I mentioned above and could be integrated into it But I thought I'd keep it as a separate feature because it's slightly separate at the same time

The case for creating a standalone vault of personal contacts data.

A transformative discovery that I made when working with LLMs last year was how valuable even tiny pieces of context data can be in increasing the personalization and specificity of model outputs.

I was backing out some hardware upgrades and doing a lot of repetitive prompting, providing the same spec and asking what would work with it.

So my "hack" / system was to simply create a markdown document containing that spec.

I could then just drag and drop that into a model to quickly provide context and guide recommendations.

I then began thinking about how this process could be scaled, so I created a few repositories and began proactively developing these "context snippets", as I call them.

Personal Knowledge Vault

I think it would be very cool to have a dedicated personal knowledge vault that the user could manage.

The only thing I've really noticed about the personal context use case is that it's important that the data is updatable. For example, I might wish to upgrade my computer specs after an upgrade and then quickly simply save that file so that it's in the system.

The user could populate the system with a rich repository of information about themselves, so long as they felt confident enough doing that:

• My address
• My secrets (as in passwords!)
• My dentist's name, address & phone number

For the purpose of context management, this personal knowledge vault could be considered one unit of data to be maintained in the RAG pipeline. And then with your new agent feature, the user could connect it to any or multiple agents.

Alternatively, as I shared before, it could be broken down into discrete areas so that the user could assign only specific parts of their personal context footprint to individual agents.

Each of these vaults and turn is just a collection of markdown files that have been sent into RAG:

Implementation could be as simple as a text editor front-end allowing the user to add, edit or delete specific entries:

They can do this only once and then connect it to an infinite number of agents.

Potential Use Cases

• Daniel and Hannah are looking for a new home. Both each have their preferences and they're trying to keep track of what they've already seen. So they create a knowledge pool for the house hunt data and update that as they go along. That knowledge pool is connected to their house hunting assistant agent.

• Daniel is looking for a new job, so he creates a knowledge pool called Job Hunt Context Data into which he populates his resume, skills and professional aspirations. He then connects that personal context data to a job hunting agent.

This could be a simple and effective but powerful way of providing a personal context data pipeline.