How Rafiq works

Every file is turned into facts and merged into one knowledge base. Text-bearing files (a Word handbook, a PDF with real text, a spreadsheet) are read directly. Photos and scanned PDFs are looked at by the model page by page, so a photo of the menu board yields the same kind of facts as a typed menu.

Big files never leave your browser whole: photos are shrunk before upload, and a PDF over 4 MB is opened in the browser first. If it has a text layer only the text is sent; if it is a scan, each page is rendered to a photo and sent in batches. The card on screen tells you what it is doing and how long it usually takes, about 20 seconds per document.

Whatever a document covers, the interview skips. The coverage board under the upload form shows which sections the documents already filled.

The interview works through twelve sections: the business, roles and shifts, the menu, equipment and procedures, opening and closing, food safety, customer service, POS and payments, house rules, safety and emergencies, escalation, and how you want onboarding paced. Rafiq always picks the next topic, going to the thinnest section first, and does not end while a required section is still short.

Each section opens with an invitation, "walk me through a normal morning", and Rafiq listens without cutting in. Then it compares what you said with what the section needs and asks for the specific missing pieces: the temperature, the exact steps, the words staff should use. "We keep the fridge cold" is never accepted as a fact; the number is.

As you talk, facts are recorded into the knowledge base live, so the coverage panel beside the conversation fills up while you speak. When the interview ends, the whole transcript is run through the extractor once more and merged, deduplicated by meaning, so saying the same thing twice does not create two facts.

Voice needs a microphone; if the browser refuses or the connection drops, the same interview continues by typing.

Each fact is one plain sentence a new employee could act on, tagged with where it came from: an interview turn, a document, or a question a worker asked that you answered. A section counts as covered when a new worker could do that part of the job from its facts alone.

Generation is manual on purpose. It takes two to four minutes and costs real model time, and you often want a second interview or a correction first. Nothing stops you generating early, but the program can only teach what the knowledge base contains.

Corrections go in the same way as everything else: type the right version into the box, or tell Rafiq in a short follow-up interview. The extractor recognises that it is the same fact and replaces the old one rather than adding a second.

The program is generated in English from the knowledge base and then translated lesson by lesson into the other languages on the business profile, so an Urdu-speaking barista and an Arabic-speaking cashier are learning the same content. Every fact in a lesson traces back to the knowledge base; the program never invents procedures.

Each lesson is a short sequence of blocks: an explanation, a quiz, and often a roleplay where the worker talks to Rafiq playing a customer or a colleague and is scored against a rubric drawn from your own facts. Roleplay dialogue runs in the language customers are served in.

Each lesson carries a tier. First-days lessons are the ones a new hire needs before their first solo shift; keep-learning lessons are released over the following weeks at the pace you chose in the interview. You can move a lesson between tiers on the program page, and that override survives regeneration.

When the knowledge base changes after generation, the program page marks the program as behind and offers to regenerate. Regenerating replaces the program, so do it before workers start rather than half way through.

A worker's link contains only a short code, not the business address, so it cannot lead them to these owner pages. Anyone who has the link can open it, so treat it like a key and make a new worker if one is lost.

The worker's language is theirs to pick, not yours: the owner's list shows "Not chosen yet" until they open the link. They can change it later from their home page.

The per-worker page shows every lesson they have done, their quiz answers, the roleplay transcripts and scores, and the questions they asked Rafiq.

These are by-products of the knowledge base, not separate work: the same facts that make the lessons are written out as documents an inspector or a new manager would expect to see. The manager's onboarding plan is the one document written from the program, day by day, in the manager's language.

Progress is computed from what workers actually did: lessons started and completed, quiz answers, roleplay scores. A module is complete when all its lessons are. Readiness is whether the first-days lessons are done.

Ask Rafiq is available to every worker from their home page. Rafiq answers only from the knowledge base and says so when the answer is not there, pointing them to the shift lead. Every question is logged here, in the worker's language and in English, so a question three people ask is a gap you can fill once for everyone.