jev-palette

A semantic command palette (⌘K) demo powered by Jev, TypeSafe AI's System One model. Type your intent — in Portuguese or English — and the palette ranks 77 commands by meaning, side by side with a classic fuzzy matcher that fails honestly on the same queries.

"jogar fora essa mensagem" → Delete message · "make it dark in here" → Toggle dark mode (100%) — while the fuzzy baseline finds nothing.

The trick

One Choice question per keystroke, with the whole command catalog as options. Jev returns the full probability distribution over all options in a single ~100ms call — the distribution is the ranking. No embeddings, no vector index, no output tokens billed.

state:     { query: "jogar fora essa mensagem" }
question:  { type: "choice", instructions: "Which command does the user want?",
             criteria: { delete_message: …, archive_message: …, ×77 } }
answer:    { choice: "delete_message", probabilities: { delete_message: 0.64, … }, confidence: 0.62 }

At ~900 input tokens per call and $0.042 per million input tokens, every keystroke costs about $0.00004.

Running it

npm install
cp .env.example .env   # put your TYPESAFE_API_KEY in it
npm run dev            # proxy on :8787 + Vite on :5173

No key yet? npm run dev:mock (or just npm run dev without a key) serves a local keyword-based stand-in so the UI works offline — a "modo mock" badge shows in the header until the real key is configured.

Layout

• server/index.mjs — tiny Node proxy; keeps the API key server-side, calls POST https://api.typesafe.ai/v1/systemone, falls back to mock mode.
• server/jev.mjs — pure helpers: request building, distribution → ranking, cost accounting. Covered by Vitest.
• src/commands.js — the 77-command catalog of a fictional mail app, shared by browser and server.
• src/fuzzy.ts — the deliberately-classic baseline matcher.
• src/App.tsx + src/useJevRank.ts — React UI: debounced (200ms), aborts in-flight calls, caches repeated queries, degrades to fuzzy if Jev is down.

Tests

npm test

Eval

50 labeled queries (eval/queries.mjs) in three slices: intent phrased in Portuguese, intent phrased in English, and name-like queries where classic fuzzy matching is strong. Run with npm run eval (live API, ~$0.004 total). Results against jev-1.13.0 (2026-09-19):

Slice	n	Jev top-1	Jev top-3	Fuzzy top-1	Fuzzy top-3
pt-intent	25	100%	100%	0%	0%
en-intent	15	93%	100%	0%	0%
name	10	100%	100%	100%	100%
overall	50	98%	100%	20%	20%

Mean Jev latency: 413ms. The single top-1 miss is "throw this away" ranked as Discard draft over Delete message — a defensible reading without more state. The name slice is the honest caveat: when the user already knows the command's name, fuzzy is just as accurate and effectively instant, which is why a production palette should run fuzzy first and fall through to Jev when fuzzy comes back empty or weak.

Schema notes (verified against jev-1.13.0)

A choice question takes instructions (the decision prompt) and criteria (a map of option id → description); the answer comes back with choice, probabilities (floats summing to 1 across every option) and confidence. rankFromAnswer in server/jev.mjs also tolerates a distribution field name, just in case.