Skip to content
Laya AI

Unofficial community documentation

Use cases

How Laya System One divides work with closed decision APIs — then where hosted prepaid keys fit if you skip local install.

When Laya fits

  • You want Apache 2.0 weights on your own machine, or a hosted prepaid key without running torch yourself.
  • The choice set is small — ticket departments, AG News–scale labels, triage templates.
  • You will fine-tune a specialist checkpoint for your workflow.
  • Local GPU latency matters, or you want a key tonight via the hosted gateway on this site.

Prefer hosted? Packs from $5, typical decide ≈ $0.0004.

Where closed APIs differ

  • You want a vendor closed API already specialized for your task, with no fine-tune step.
  • One choice question needs a very wide label inventory in a single prompt.
  • You lean on soft / full probability distributions as a first-class product surface.

Published head-to-head numbers (including wide-label and baseline notes) live on Benchmarks. Still want to try hosted Laya? From $5.

Ticket routing

Good fit

  • A support message needs a small department choice, an ordinal urgency score, and yes/no flags such as churn or an explicit refund request. That is the README Router example and the triage preset laya.triage_questions().
  • choice, score, and yes/no run in one forward pass. The README shows a confidence gate at 0.85 for automatic routing versus human review. Fit temperatures on your own data before trusting that threshold.

Poor fit

  • A very wide support queue with dozens of teams in one choice question — keep label sets small, or shortlist first.
  • Writing the reply. Laya does not generate text; pair with a drafting model only if you add one yourself.

Content safety

Good fit

  • Spam / phishing style gates and prompt-injection style yes/no flags in a single decide call — see the moderation and guard presets in the README.
  • Hosted templates on this site (prompt-guard, content-moderate, scam-spot) share the same prepaid balance as /v1/decide.

Poor fit

  • A safety stack that only checks English-checkpoint confidence on text the model cannot read — route multilingual traffic to the multilingual checkpoint first.
  • Sole control for jailbreak / toxicity: treat model scores as a signal, not the only gate.

Intent classification

Good fit

  • English intent with a modest label set, or other languages when Router sends the request to laya-multilingual.
  • Small choice sets that match the ticket / AG News style examples in the README.

Poor fit

  • Non-Latin text on the English checkpoint without routing — confidence can stay high while accuracy collapses.
  • One choice question with a very large option list at default head budget — raise head_max_len or shortlist first.

Scoring

Good fit

  • A short ordinal rubric: the README score primitive returns an expected level, a distribution, and a confidence.
  • After domain fine-tuning when your rubric matches the specialist checkpoint.

Poor fit

  • Fine-grained sentiment such as SST-5 — the README treats ordinal score as the weakest primitive.
  • Expecting a base install to match a fine-tuned specialist score out of the box — see /benchmarks/ for which checkpoint owns which figure.