Optional AI adapter

A model may propose.
It never gets the final word.

The public lab can optionally call a model through a provider-specific server adapter, while Orbweaver core stays provider-neutral and every result follows the same semantic review path.

Authoritative source on GitHub ↗

Model-assisted generation is an opt-in website capability—not a dependency of the Orbweaver package. Deterministic generation remains the default; the model path is disabled until the site operator enables it.

GraphProposal v1 · structured JSON · explicit human acceptance
Intent + optional evidence summaryonly user-supplied, non-sensitive inputCloudflare Workers AI JSON modeprovider-specific server boundaryServer and browser validationuntrusted proposal; never drawing instructionsInspect, accept, exportvisible actions reserved for the person
01

The contract given to the model

The adapter requests a GraphProposal v1 through JSON Schema. It asks for semantic entities, relationships, groups, reviewable evidence, claims, warnings, and annotations where justified.

Meaning, never geometry

Coordinates, positioned scenes, SVG, HTML, CSS, renderer configuration, executable markup, and invented source provenance are prohibited. The host overwrites generation metadata with the actual provider and model identity.

02

Cost and operating boundaries

  • Deterministic mode costs nothing and makes no provider call.
  • Model mode estimates USD 0.001 and caps one request at USD 0.002; the site displays both values with the result.
  • Requests are limited to 16 KiB, time out after five seconds, and can be cancelled in the browser.
  • Prompts, evidence, proposals, and acceptance decisions are not stored by Orbweaver.

Never submit secrets, personal information, confidential records, or regulated data. A valid proposal is contract-conforming—not factually proven.

03

Why it stays optional

Package

Deterministic by default

No AI SDK, account, key, provider, prompt, or network dependency belongs in @semanticintent/orbweaver.

Website host

Owns the provider choice

The host configures the model binding, spending boundary, availability, provider terms, and any future rate controls.

Open the AI Proposal Lab →