I have also open sourced the skills we used in Sentry to prove out this latest iteration.
https://t.co/CNtfq4ZqYs
Please use it responsibly. If you find something that others have missed, validate it, and send something up to bounty programs.
p.s. Mythos is FUD
For the final refine phase, we implemented a cache-optimized Product Quantization (PQ) layout specifically tailored for late interaction.
Evaluated on ColBERTv2.0 embeddings, it results in 10 ms single-CPU retrieval on large-scale datasets (MS MARCO-v1, LoTTE Pooled).
@ibragim_bad@Shevan05@agolubev13 What I find most interesting is that you found a way to make Gemini 3.1 Pro work when everyone has written it off as useless. I don't know of any other harnesses that are able to make it useful.