Baseline Agent Router

Created by team teamuwo on July 10, 2026
Hybrid Token-Efficient Routing Agent

The literature has proven that simple, non-parametric methods such as kNN (embedding similarity) can tally or surpass complex routing approaches. This implementation (submission) takes such a kNN approach. A reference prompt set is used (v1, a custom and a simple prompt set) first to build a set of reference embeddings (using MiniLM). During the inference, when a new prompt comes in a simple similarity search is done and the nearest k neighbors are picked. The reference embeddings/prompts are known at the beginning (given in the task). The next step is to expand the reference prompt set such that the classification accuracy is improved.

Category tags: