An Update on CueCrux
A slightly more informal January update on CueCrux progress toward the MVP.
Hello, and a slightly more informal update this time.
CueCrux was officially set up in December, and we’re now properly into the swing of things. That mostly means a mix of practical work: getting an office sorted, tightening up systems, and pushing towards the first MVP.
If all goes to plan, that MVP should be ready for release later this month.
Early testing has been very encouraging. The core engine is behaving as expected, which is always a relief, and we’ve made solid progress on the embedding side of the platform. In particular, we’ve spent time polishing and upgrading our approach to vector embeddings, and we’re now seeing multi-lane embeddings running cleanly across the system.
At the moment, artefacts are stored in PostgreSQL, which has been a reliable backbone while things are coming together. Alongside that, we’re developing our own artefact database designed to operate at a very different scale. The aim is a system capable of holding billions of artefacts and retrieving them in a significantly reduced time period, effectively close to instantaneous access even as the knowledge base grows.
Longer term, that database will likely run side by side with PostgreSQL rather than replacing it outright. Redundancy, traceability, and fallback matter more than elegance when you’re dealing with knowledge you expect people to rely on.
A lot of work this month has also gone into how artefacts relate to one another. We’ve been strengthening the interrelationships between artefacts so that, later on, the system has something richer to reason over than isolated chunks of information. That work feeds directly into how answers are formed and how evidence can be interrogated.
One notable shift we’ve made is moving more of that relationship processing before search rather than after it. This makes the embedding stage more involved, but the payoff has been clear. Results are better grounded, references are stronger, and the system is able to reason about context more effectively than when everything is done on demand. In reality, it will probably settle into a hybrid of both approaches over time.
As the platform gets closer to release, we’ll be opening up a short list of early evaluators. Those first users will matter a lot. Not just for feedback, but because real usage always reveals things no amount of internal testing ever does. I’m genuinely looking forward to seeing the kinds of use cases people push into the system once they have their hands on it.
We’ve spent a long time thinking about who might benefit most from CueCrux and how it should behave. But at some point, you have to put it in front of people and let reality have a say.
This feels like a good place to start.
More updates soon.