bev-ai

Thinking, Slow and Slower. The slowest classifier in the world, at Jev prices.

Bev uses Jev under the hood, so she is exactly as cheap and exactly as accurate as Jev. But after getting the real answer in under a second, she performs minutes of visible "reasoning" before handing it over. Every result tells you how long the real thinking took.

$ npx bev-ai "I was charged twice for my subscription" --effort low

  Bev · System Three
  Ticket:  "I was charged twice for my subscription"
  Options: billing · shipping · technical · other
  Effort:  Low (about 10s)

  Thinking... (press r to rush, Ctrl+C to leave)
  [0:01] Bev has taken a number for your ticket. It is number 4,512.
  [0:04] Bev put it in the wrong pile. Bev is starting that part over.
  [0:07] What about technical? Bev is considering it. Bev is done considering it.
  [0:09] Decision made: billing. Bev is stamping it.

  Verdict  BILLING
  Bev is 97% confident. (Jev's actual confidence: 88%)
  Bev thought for 9s. Actual thinking time: 0.38s.

Not affiliated with TypeSafe AI. Bev just works here.

CLI

Requires Node 22 or later.

npx bev-ai "<text>"                                   # support queues: billing, shipping, technical, other
npx bev-ai "<text>" -o "yes,no,maybe"                 # your own 2 to 8 options
npx bev-ai "Great news, I got promoted" -p vibe       # good news, bad news, neutral
npx bev-ai "<text>" -e bev                            # low ~10s, medium ~30s (default), high ~90s, bev 3 to 5 min
npx bev-ai review "Per my last email..."              # fine / passive-aggressive / furious / will get you fired (10 minutes)
echo "<text>" | npx bev-ai --json > verdict.json      # JSON on stdout, Bev's thinking on stderr

While Bev thinks, press r to rush her. Each rush adds 10 seconds. Bev does not appreciate being rushed.

On Fridays Bev is out of office and Gary covers. Gary takes about a third of the time and has seen worse. --gary gets him any day.

Library

import { classify, startClassify, review } from "bev-ai";

const result = await classify("My package never arrived", ["billing", "shipping", "technical", "other"], {
  effort: "low",
  onThought: (t) => console.log(t.text),
});

result.choice;          // "shipping", straight from Jev
result.probabilities;   // Jev's real probabilities
result.bevConfidence;   // 0.97, always
result.jevConfidence;   // Jev's actual confidence
result.bevMs;           // how long Bev took
result.actualMs;        // how long Jev took

startClassify and startReview return { result, rush } so you can rush Bev from your own UI. Options can be an array of labels or an object of { label: description }. Pass signal to stop waiting. Errors are BevErrors with messages in Bev's voice.

Jev keys

Bev needs a Jev key for real answers. Set one; the first one found wins:

| Variable | Backend | |---|---| | TYPESAFE_API_KEY | TypeSafe direct | | AI_GATEWAY_API_KEY | Vercel AI Gateway (typesafe-ai/jev) | | OPENROUTER_API_KEY | OpenRouter (typesafe/jev-1.13) |

With no key, Bev runs in mock mode: a keyword heuristic stands in for Jev, and the CLI says so. Your text is only sent to the backend you configured, and Bev stores nothing.

Why is Bev slow?

People value work more when they can see the effort behind it (Buell and Norton, The Labor Illusion, 2011). Years of reasoning models taught all of us that waiting means intelligence. Bev sells the wait people have learned to trust, and then tells you the truth.

License

MIT