Article

What to prepare
before adding an assistant

Four things decide how well an embedded assistant serves your users: your documentation, your API description, how you identify users and how you test. Here is what to get ready for each.

By the Orrin team · 2 min read

Most of what makes an embedded assistant good happens before anyone asks it anything. If you build property management software and are thinking about adding one, these are the four things to prepare, in the order Orrin’s console asks for them.

Documentation your users would recognise

The assistant answers from your documentation and nothing else, so its answers are only as complete as your help center. Three habits help most: one topic per article, clear headings, and the words your users actually use. If your users say “late fee” and your articles say “delinquency charge”, add the plain term.

Orrin reads public help pages, pasted text, and Markdown, HTML, text, JSON and CSV files.

An OpenAPI file for the tasks you want

To prepare actions, Orrin reads an OpenAPI 3 or Swagger 2 file in JSON. If yours is in YAML, convert it to JSON first.

Give every operation a clear summary and description, because that wording is what a request is matched against, and mark required parameters as required. Start with a handful of everyday tasks rather than the whole API: you choose what the assistant can reach, and nothing is exposed until you switch it on.

A way to say who the user is

An assistant inside your product must only ever do what the signed in user is allowed to do. In Orrin, your backend signs a short lived token for each user with a key you generate in the console. The token carries the user’s id, their company, their role and their scopes.

Orrin keeps only the public half of the key; the private key is shown once, at generation, and never stored. Test tokens live ten minutes, and you also list the web addresses your product runs on.

A plan for testing

Before your users see it, ask it what your support team is asked most, in your users’ words, and open each answer’s citation. Then read what it could not answer: that list is your documentation backlog, and filling it makes the next round better.

Try action requests too, and check that each preview matched the right operation at the right level and read the right values.

No AI team required

None of this needs an AI team. It needs a help center you keep up to date, the API you already have, and a clear rule for who may do what. The console walks through the same four steps, and the Sandbox plan is free.