CONTACT
Call Us: 1300 179 987
LEGAL
ABN: 81 632 590 759
Neural RAG Router — routes every query to the right-sized model.
A neural RAG router that sends every query to the cheapest model that can answer it well.
Most AI systems push every question through one expensive model. Cortex reads each query first, predicts how hard it is, and routes it — simple questions to a small, fast model; complex, multi-step reasoning to a heavyweight. The decision takes under five milliseconds.
Because the router is learned rather than rule-based, it holds near-heavyweight accuracy while cutting inference cost by 60–90%. And it explains every routing decision in plain language, so nothing is a black box. Run it on your own GPUs with vLLM or Ollama, or against the cloud with Anthropic Claude — no lock-in.
A multi-head transformer gauges each query's complexity and the cost and accuracy every model would deliver.
It sends the query to the lightweight, medium, or heavy model — whichever wins on accuracy, cost, and latency.
Every decision comes with a plain-language reason and a confidence score. No black box.
The router keeps improving from real outcomes through reinforcement learning.