codecoradev/laya-service repository overview

⁠laya-service

Typed decisions over any state in one forward pass — an HTTP sidecar for the Laya⁠ decision engine (Apache-2.0, 27k★).

~0.2–0.6 s per decision · ~170 calls/min · CPU-only · stateless · Apache-2.0

Jev-wire-compatible — official TypeSafe Jev clients work against this service by swapping the base URL and bearer token only (POST /v1/systemone, optional model accepting jev-latest). Ceiling differences: choices ≤ 20 (Jev: 255), state context ~1k tokens (Jev: 32k).

The service is stateless and policy-free: it only answers typed questions (choice / score / noul) over any state you send. What to do with an answer — auto-act, log, escalate — always lives in your caller.

• Typed answers, not free text — request schemas are validated before the model is touched (fail-fast errors)
• Calibrated confidence (RLCD) — common policy: auto-act at ≥ 0.85, escalate below
• Nothing shared ships inside the image — the service is fail-closed; you bring your own auth at the edge

Built and maintained by CodeCora⁠ — pragmatic AI infrastructure.

⁠Quick start (~30 s)

docker run -d --name laya \
  -p 8080:8080 \
  -v laya-hf-cache:/data/hf \
  --memory 4g \
  codecoradev/laya-service:latest

Weights (~644 MB) download from Hugging Face on first start and are cached in the volume. Watch readiness — status: "ok" means the model is loaded:

curl http://localhost:8080/healthz

First run downloads ~644 MB of weights from Hugging Face (~1–2 min on a decent connection, then cached). Fresh named volumes work out of the box — the image pre-creates /data/hf with correct ownership. Using a bind mount instead? Pre-chown it to the service user (uid 999):

mkdir -p laya-hf && sudo chown -R 999:999 laya-hf   # then -v "$PWD/laya-hf":/data/hf

A volume created by an older (pre-2026-10-05) image is root-owned; fix once with docker volume rm laya-hf-cache (it just re-downloads) or chown it.

Environment: LAYA_CHECKPOINT (default multilingual), LAYA_DEVICE (default cpu), AUTO_THRESHOLD (default 0.85). Multi-arch: amd64 + arm64.

⁠Ask a question

Four equivalent endpoints, one contract: /v1/systemone (TypeSafe-canonical), /api/v1/systemone, /api/v1/decide, /v1/decide (legacy).

curl -X POST http://localhost:8080/v1/systemone \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "jev-latest",
    "state": {"body": "I was charged twice, refund now or I cancel!"},
    "questions": {
      "intent":   {"type": "choice", "instructions": "What does the customer want?",
                   "criteria": {"refund": "money back", "technical": "bug", "shipping": "delivery"}},
      "urgency":  {"type": "score",  "instructions": "How urgent?",
                   "criteria": ["routine", "soon", "critical"]},
      "escalate": {"type": "noul",   "instructions": "Needs a human now?"}
    }
  }'

model is optional — omit it (or send latest / jev-latest / jev-preview) to use the preloaded checkpoint. Real response (2026-10-05, live):

{
  "model": "multilingual",
  "answers": {
    "intent":   {"type": "choice", "choice": "refund",
                 "probabilities": {"refund": 1.0, "technical": 0.0, "shipping": 0.0},
                 "confidence": 1.0},
    "urgency":  {"type": "score", "score": 1.92,
                 "legend": {"0": "routine", "1": "soon", "2": "critical"},
                 "probabilities": {"0": 0.005, "1": 0.065, "2": 0.93},
                 "confidence": 0.75},
    "escalate": {"type": "noul", "noul": 0.46}
  },
  "usage": {"input_tokens": 123, "output_tokens": 0}
}

usage.input_tokens is a chars÷4 estimate; output_tokens is always 0 (the model is non-autoregressive — no generated text).

⁠Question types

state accepts a plain string, an email/ticket-like object, or a JSON document — passed to the engine untouched. Keep it short: context is 1024 tokens (a ~2.5k-token state takes ~17 s vs ~0.3 s at normal sizes).

⁠Endpoints & errors

Errors are Jev-style: {"error": "…", "detail": …}. 422 malformed body (missing state/questions, bad types — Jev parity) · 400 semantically invalid (unknown model, wrong criteria shape) · 401 wrong bearer (if you add auth) · 500 prediction failure · 503 server busy (all run slots taken, queue full) with a Retry-After header: retry the same request.

⁠Measured performance (2026-09/10, real E2E)

confidence is calibrated (RLCD) — treat it as meaningful. Sweet spots: CS/ticket triage (intent + urgency + needs-human), event triage, spam/phishing gates, small model routing. Not for: long-context reasoning, >20-way choices, content moderation.

⁠Tags

⁠Links

• CodeCora — the team behind this: https://codecora.dev⁠
• Upstream Laya (Apache-2.0 by Convai Innovations): https://github.com/NandhaKishorM/laya⁠

⁠More from CodeCora

• Uteke⁠ — the remember-side of the same architecture: local-first semantic memory for AI agents (this service is the decide-side). Single Rust binary, ~45 ms recall, no cloud required.
• Cora Code⁠ — CLI-first AI code review (BYOK, pre-commit hooks, CI) — the reviewer that watches this very repo's pull requests.

Apache-2.0. Laya is Apache-2.0 by Convai Innovations — this service only wraps it.