Integration

Make the boundaries explicit.

SuperER connects language to application data. Your application owns the catalog, domain rules, and authority to act.

1. Define your catalog

Use stable canonical IDs. Attach aliases and the attributes tools consume, such as team IDs or city and state. Version the catalog so stale references can be detected. Keep user-visible result order tied to the artifact snapshot that produced it.

2. Interpret the request

Use a deterministic parser for a bounded command language, a model adapter for broader language, or a combination. Return typed mentions, roles, constraints, and state operations. Unsupported language must remain explicit; an omitted constraint must not silently become “no preference.”

3. Declare each tool’s contract

Map tool fields to evidence dimensions and projections. Declare required fields, validators, and relevant consequences. A routing tool may need a team; a notification tool needs an identity and contact address. The contract determines which ambiguities must be resolved.

from superer import FieldRule, ToolContract

notify = ToolContract("notify_person", (
    FieldRule("person_id", "person"),
    FieldRule("email", "person", "email"),
))

4. Maintain a session

Use revision checks when submitting a turn or accepting a clarification. Persist structured questions and their candidate options alongside the conversation. Do not reconstruct a pending selection from assistant prose.

5. Validate at dispatch

Before execution, check permissions and the freshness of the catalog, artifact, and memory evidence. Keep business-action retries idempotent. SuperER’s proposed parameters do not grant permission to send a message, make a booking, or change an account.

Evaluate the whole path

Measure candidate recall, identity and role accuracy, state updates, wrong tool parameters, and unnecessary clarification. Include slang, corrections, long conversations, missing candidates, and model failures. Report completion and abstention together.

The optional Multivon integration can score recorded proposed actions. Executed outcomes require separate observations from the application.