Relaiquary

Overview & Motivation

Humans are creatures of stories, and our lives are the stories we tell ourselves. There is something satisfying about putting pen to paper, or in this case, the conscious flow of thought into keystrokes. To remain unfiltered and wholly authentic though, it is imperative that the contents are kept secure and shielded from the view of others.

What if this story, this data, could be seen without risking the author's filter? The answer seemed simple to me: Locally run and open source Large Language Models (LLM).

A machine trained on the stories of humanity with no hidden agenda of its own.

Not a co-host of your story. Not a friend. An impartial* judge and analytics engine to spot your shortcomings and triumphs when you unknowingly gloss over them yourself.

This is what I wanted to create, and I think I've succeeded in my goal.

Relaiquary is a self-hosted application where the user can write, edit, manage, and store their most precious memories and thoughts, their relics, in plain markdown files to remain application agnostic and future-proof. These are then linked to a SQLite database, ready to be visualized into a beautiful map of your psyche. You may then query the oracle (the LLM), or read the divinations (analytics) made on your latest journal entry, your past week, and even your past month.

While you should be able to hook up and do inference with a cloud provider's hardware (which I don't recommend), your own hardware and inference capabilities will determine the usefulness, and viability, of the context-heavy tasks like month-long divinations. Everyone with a decent laptop that has a compatible graphics card and operating system should at least be able to get database visualization, basic queries, and single-entry analytics working, assuming a model that fits your hardware limitations.

The obvious (but maybe unhelpful) recommendation is that this application is run on a machine of your choice with the models themselves running on your own privately owned GPU-cluster to never have to worry about the context length and VRAM demands. But how many can swing that in this economy?

It is currently in active private development and being prepared for an initial open-source release on GitHub.

Key Highlights

• Local-First Architecture: Operates entirely offline, with your own hardware, and your own models.
• Application agnostic files: No strange or proprietary formats. Do as you will with your own generated markdown files.
• Free and Open Source: Apache 2.0 License (I need to check this to make sure there are no viral dependencies).

*All models have inherent biases from training. Critical thinking is still, unfortunately, advised.