Semantic Router alternative: decision models for routing decisions

What semantic-router solves

semantic-router (aurelio-labs) matches user utterances against predefined routes in vector space and returns a routing decision in milliseconds, skipping LLM generation entirely. It has become one of the most popular routing layers in LLM stacks, and Red Hat’s vLLM Semantic Router brought the idea into production inference stacks.

Its sweet spot is speed: embedding lookup is cheap and fast. Its trade-off is that routes are semantic-similarity clusters — they answer "which route is closest", not "which option is correct given this state and these criteria".

Where decision models differ

System One decision models take structured state plus typed questions (yes/no, choice, score) and return calibrated judgments: a probability, a distribution over candidates, or an ordinal score. Routing becomes a judgment with explicit criteria rather than a nearest-neighbor match — closer to "route by policy" than "route by similarity".

The two compose well: semantic-router for ultra-cheap first-pass bucketing, decision models for the ambiguous middle where criteria matter. If your routing rules are policy-like ("premium users with >30k requests go to the dedicated pool"), a decision model expresses them directly in the question.

Try it in one endpoint

EasyJev serves the whole decision-model catalog behind a single typesafe-sdk-compatible endpoint with free output tokens — routing decisions never pay for generation. Swap base_url and api_key, then pick any model per request.