hermes-laya

Laya System-1 typed decisions as a Hermes Agent plugin.

Laya classifies text against typed questions in one local CPU forward pass with calibrated probabilities. It never generates text — it is the fast, deterministic first read (triage, routing, urgency, risk flags) before the LLM does any actual reasoning.

Tools

laya_triage

Pre-built message triage, one call:

{
  "text": "Pasien pingsan di ruang tunggu, tolong sekarang!"
}

Returns:

{
  "success": true,
  "department": "ops",
  "department_confidence": 0.99,
  "department_probabilities": {"billing": 0.0, "ops": 0.99, "clinical": 0.01, "technical": 0.0},
  "urgency_score": 1.95,
  "urgency_label": "emergency",
  "escalation_probability": 0.263,
  "model": "multilingual",
  "latency_ms": 412,
  "usage": {"input_tokens": 210, "truncated": false}
}

Custom routing targets via the optional departments object (label -> gloss).

laya_predict

Full typed-decision surface — batch any mix of question types in one pass:

{
  "state": "We were billed twice for March. Refund the duplicate today.",
  "questions": {
    "intent":    {"type": "choice", "instructions": "What does the customer want?",
                  "criteria": {"refund": "money returned or duplicate charge reversed",
                               "technical_help": "a bug or outage",
                               "question": "general question"}},
    "urgency":   {"type": "score", "instructions": "How urgent?",
                  "criteria": ["can wait", "today", "blocking"]},
    "is_threat": {"type": "noul", "instructions": "Does the customer threaten to cancel?"}
  },
  "min_confidence": 0.8
}

type	criteria	returns
choice	dict label -> gloss, or list	choice + probabilities
score	ordered list, low -> high	score = expected level (float)
noul	optional {false, true} glosses	noul = calibrated P(true)

Options: model (multilingual default — use it for non-English text), max_len (long documents), min_confidence (abstain below threshold).

Install

hermes plugins install agusbyna/hermes-laya
hermes plugins enable hermes-laya

The install flow asks for consent before pulling the Python dependencies (laya, which brings PyTorch — expect a ~2 GB download on first enable). On first tool call the checkpoint downloads from Hugging Face (~1.5 GB, cached afterwards).

Verify:

hermes plugins doctor hermes-laya --ci
hermes plugins list

Performance & resources (measured)

• First call per process: ~15-30 s one-time checkpoint load. Subsequent calls: ~0.3-0.7 s per short message on CPU.
• Resident memory: ~2 GB once the model is loaded.
• Outputs are fully deterministic; question order does not affect results.
• Every Router build does a Hugging Face revision check. With both checkpoints cached you can set HF_HUB_OFFLINE=1 in ~/.hermes/.env to skip it (verified working).

Config

Set via hermes config set plugins.entries.hermes-laya.config.<key> <value> or the Desktop plugins tab:

key	default	meaning
default_model	multilingual	checkpoint (multilingual / english / typed-decisions)
max_len	0	state token budget; 0 = checkpoint default (1024 multilingual / 512 english)

Bundled skill

hermes-laya:laya-system1 documents the question types, the measured rubric behavior (over-triage bias, noul vs score, calibration limits), and the operational pitfalls. Load it with skill_view("hermes-laya:laya-system1").

Limits (be honest with yourself)

• This is System 1: right for triage, routing, urgency, yes/no risk flags; wrong for anything needing reasoning over the text.
• score levels are not calibrated thresholds — the triage rubric is example-anchored and deliberately biased toward over-triage (in testing it never under-flagged an emergency, but routine items can land at soon). Read probabilities, not just labels.
• Department routing on Indonesian tickets measured 4/8 correct with gloss tuning; the checkpoint is the bottleneck, not the rubric.

Credits & license

• Laya by Receptron — Apache-2.0.
• Hermes Agent by Nous Research.
• This plugin: Apache-2.0.