I cut cost and latency on my search engine with Jev [D]

I built indiedex.gg, a hidden-gems game search engine on steam data. Search is the front door: type something like “a game like Hades that feels cozier and is co-op,” and it should actually understand that mix of reference + vibe + filters.

Those compound queries used to go through a DeepSeek extraction call that turned the sentence into structured pieces (reference game, filters, tags, vibe). It worked, but it was slow and expensive on the hot path.

I swapped that step to TypeSafe Jev on OpenRouter’s Decisions API. Jev doesn’t write a free-form answer. It answers a fixed set of closed questions in parallel (yes/no odds, choices), and I compose that with my existing regex helpers into the same extraction shape I already had. Same search pipeline after that, just much faster routing.

Bakeoff on 50 cases:

So extract got roughly 10× faster, end-to-end route roughly halved, cost dropped a lot (around 6x), and quality held on my fixture set.

If you’ve been building “NL in → structured intent out → tools do the work” instead of a chatbot, this pattern felt like a better fit than another chat completion.

I think models like Jev will open the doors to a lot of cool UX features, decision engines, smart filtering, auto moderation and different ways to interface with apps and machines.

Happy to answer questions about the hybrid setup.