1. Support ticket routing
Classify an incoming message into billing, technical support, sales or another queue; detect explicit refund requests and estimate urgency in the same call. The output can route clear cases and send ambiguous ones to manual triage.
Good fit
Few distinct queues, labelled historical tickets, concrete queue descriptions and a human escalation path.
Watch for
Messages with multiple intents, new queues, non-English text sent to the English checkpoint and pressure to automate from an unvalidated confidence score.
2. Explicit fact detection
Noul questions are useful for facts visible in the state: “Does the user request a refund?”, “Does the message mention an account takeover?” or “Is a cancellation threat explicit?” Several factual answers can feed deterministic business rules.
Avoid asking whether a customer is “good,” a message is “appropriate,” or an action is “safe” without a precise rubric. Broad judgments hide multiple questions and produce harder-to-calibrate probabilities.
3. Model and tool routing
Before calling an expensive model, classify a request by domain or required capability. Straightforward cases can go to a small model while coding, legal or deeply ambiguous requests go to a frontier model or specialist workflow.
Inspect the request
Ask narrow questions about domain, sensitivity and task type.
Apply code-owned rules
Use the typed outputs as features; keep budgets and permissions in deterministic code.
Measure downstream success
Routing accuracy alone is not enough—track whether the selected route actually resolves the task.
4. Prompt guardrails
Laya can screen prompts for observable jailbreak or injection patterns before an LLM call. The project’s held-out guardrail results are around 0.71–0.76 rather than near-perfect, so it should be one layer among input normalization, permissions, tool constraints and output checks.
Do not treat a single classifier as a security boundary. False negatives can pass a malicious prompt; false positives can block legitimate work.
5. RAG passage relevance
Given a query and retrieved passage, estimate whether the passage contains evidence relevant to the question. This can filter weak chunks before generation or help choose between retrieval strategies.
The project’s best listed result on its application suite is 0.657. That is enough to explore reranking or gating, not enough to assume reliable factual verification without task-specific evaluation.
6. Email spam and phishing triage
Laya reports very high results on spam and phishing suites, but both were represented in training. This makes them demonstrations of learned task fit rather than proof of broad zero-shot detection. Retrain and test on your organization’s current message distribution, including newly emerging attack patterns.
7. Local-first cascades
Use a specialized Laya checkpoint for clear, frequent cases and escalate the uncertain remainder to Jev, a general LLM or a human. This is often more realistic than asking one model to maximize accuracy, latency, privacy and cost simultaneously.
The threshold is a product decision
A threshold changes automation coverage and error cost. Choose it from validation curves for each task, then monitor drift. A single global threshold across choice, score and noul is unlikely to be optimal.
Poor fits
| Task | Why Laya is a poor fit | Use instead |
|---|---|---|
| Write a customer reply | No text generation | A generative LLM after routing |
| Summarize a document | Answer space is open-ended | A summarization model or LLM |
| Select from hundreds of verbose labels | Default option-token budget collapses | Jev, hierarchical routing or retrieval shortlist |
| High-stakes final approval | Calibration and distribution shift remain task-specific | Decision support plus deterministic checks and human review |
| Explain a decision | No generated rationale | Store evidence separately or use an explanation workflow |
Before implementation
- Define the exact decision and enumerate acceptable outputs.
- Collect representative labelled examples, including ambiguity and “none of the above.”
- Choose the checkpoint deliberately; do not rely on confidence to detect a language mismatch.
- Measure confident errors and per-class behavior.
- Decide what happens when confidence is low or the input is out of scope.
- Keep irreversible side effects behind deterministic permissions and review.
Evidence source: Laya’s consolidated benchmarks. Interface and preset examples: Laya repository. Recommendations are independent interpretations, not claims by the maintainers.