ollex 2.0.1

Open LLM Experimental Workbench

Ollex (Open LLM Experimental Workbench)

Ollex is a little desktop experiment built to get a bit of hands-on experience fiddling with this AI caper. It's a very basic, naive implementation but I knew nothing at the time. The uncharitable amongst you will suggest that situation hasn't changed. So it goes.

Genesis

Ollex was developed at Origin Energy as part of the Jindabyne initiative. While not part of our core IP, it proved valuable internally, and we're sharing it in the hope it's useful to others.

Kudos to Origin for fostering a culture that empowers its people to build complex technology solutions in-house.

Premise

The use case was simple ...

Given a reasonable sized source document, load it up so we can ask questions about its contents.

Turned out to be more fiddly than expected. The final result was a simple CLI tool with a couple of subcommands:

• olx db: Create / extend a ChromaDB vector database with embeddings from one or more Markdown formatted documents.
• olx query: A simple tool to run queries using one or more LLM models against a ChromaDB database that has been created using olx db. This can be run, either in batch mode, or as an iterative REPL.

The intent behind the olx query tool is this ... given:

• n different embeddings of a given source document
• m LLM models
• p system prompts

... process queries against the n * m * p possible combinations of these to see how each performs.

TL;DR Every one of them is a lottery.

Installation and Usage

See Ollex on GitHub.