Local quickstart

Requires Node.js 20+, about 2–3 GB of free RAM, and a one-time ~1.7 GB model download for native Laya.

Install the CLI

npm install --global @banh/cli@alpha
banh --version   # 0.1.0-alpha.1

Pin the exact version with @banh/cli@0.1.0-alpha.1. APIs and syntax may change during alpha.

Write a text workflow

Save as urgent-routing.yaml:

version: 1
process: urgent_routing
input:
  type: text
decisions:
  urgent:
    decide: whether
    question: Does this need immediate attention?
flow:
  - when: urgent.probability >= 0.85
    do: return
    value:
      route: escalation
  - else:
      do: return
      value:
        route: support

banh validate urgent-routing.yaml
banh run urgent-routing.yaml --text 'Our checkout is down.' --json

Output is {"route":"escalation"} when P(urgent) ≥ 0.85, otherwise {"route":"support"}. The threshold is your policy, not a prediction guarantee.

• validate compiles without loading a model.
• The first run downloads Laya from Hugging Face to ~/.cache/receptron-laya (override with LAYA_CACHE or --cache-dir). Inference runs on CPU; no Python or model server needed. Verified on Linux; macOS and Windows are unverified.
• --json prints process, output, decisions and timing. --verbose sends runtime events to stderr.

Run it with Jev

Set BANH_INFERENCE_TOKEN to your TypeSafe API key, then:

banh run urgent-routing.yaml --text 'Our checkout is down.' --model jev --json

Decisions run on the TypeSafe API; rules still run locally. No model download needed. See inference providers to pin a model.

Next: workflow syntax, decisions, examples.