Routing with Laya

A router runs before every request, so its own latency and cost land on the critical path. Laya answers a routing question in one forward pass with no generated tokens, which makes it a natural fit for choosing an LLM tier, a support queue or a tool.

0 projects

Nothing matches yet. Submit a project

Built-in router preset

The core package ships a question set for small-versus-frontier model routing:

import laya

agent = laya.load("convaiinnovations/laya")
routing = agent.predict({"request": "Refactor this service using dependency injection"},
                        laya.router_questions())

For LangGraph there is a conditional-edge router with a confidence fallback:

from laya.integrations.langchain import LayaRouter

router = LayaRouter(
    criteria={"billing": "invoices, charges", "tech": "bugs, outages"},
    confidence_threshold=0.80,
    fallback="human_agent",
)
workflow.add_conditional_edges("triage", router)

Note that laya.Router is a different thing: it picks which Laya checkpoint answers, based on script and language.

Measured: Laya vs a reasoning model

laya_router put Laya and GPT-5 nano behind one endpoint and tested 180 labelled requests. Both scored 0.600 route accuracy, at 184 ms p50 for Laya and 6,415 ms for GPT-5 nano. The README also found that rewording the tier descriptions moved accuracy by 21 points.

git clone https://github.com/glukicov/laya_router && cd laya_router
uv sync --all-extras
uv run laya-router serve --backend laya

Projects

  • system-one-router: an OpenAI- and Anthropic-compatible Go gateway where Laya or Jev answers topic, complexity, risk and private-data questions, then picks the cheapest adequate model.
  • pi-pignon: a pi coding-agent extension that picks a model tier per prompt from a local laya-serve (about 75 ms on Apple Silicon, per its README) or Jev.
  • laya-mcp (NVentimiglia): a route_task MCP tool for light vs frontier routing.

Caveats

In laya_router, Laya got 97% of small requests right but only 7 of 56 medium ones, so check each tier separately, not just overall accuracy. Write tier descriptions with concrete examples, and send low-confidence answers to a fallback.

More ways to use Laya