An AI assistant only knows what your documents teach it. Returns, hours, warranties: what to include, how to write it so it's clearly understood, the traps to avoid, a way to test before going live and the habits that keep the knowledge base up to date.
Isaac Poirier
Founder, Groupe Echo Inc.
An AI assistant connected to your website knows nothing about your business. It doesn't know your opening hours, your return policy or how long the wait is to get a hedge trimmed in June. Every accurate answer it gives will come from the documents in its knowledge base. And the wrong answers will very often come from those same documents: a missing piece of information, an outdated page, two files that contradict each other.
The method comes down to five steps: start from your customers' real questions, upload the documents that answer them first, write one question per section with its complete answer, keep a single source for each topic, then test with real questions before going live. It doesn't take any technical skills: it's a writing job.
With the questions. The instinct is to open the "Policies" folder and upload everything. The result: a large knowledge base that does a poor job of answering real questions.
For a week or two, write down the questions your customers ask by email, over the phone, in store and on social media, in their own words. A handful of topics will keep coming up: those are the ones to cover first.
Along the way, you may run into one of two findings. If your policies aren't written down anywhere, start by writing them: no AI can give accurate answers based on vague rules, and this work helps your team too. If you get few questions, or always the same ones, a good FAQ may be enough: our article on choosing between an FAQ and an AI assistant will help you decide.
For a small business, the list often looks something like this:
| Document | Questions it should answer |
|---|---|
| Hours and contact information | Are you open on Easter Monday? Where can I park? |
| Delivery and pickup | Do you deliver to Laval? How long does it take? |
| Returns and refunds | Can I return a sale item? Without a receipt? |
| Warranties | Is the frame covered? Who pays for labour? |
| Service rates | How much is a tune-up? Is the quote free? |
| Appointments | Do I need to book? How do I cancel? |
| Service area and season | Do you come out to Mirabel? When does mowing start? |
A bike shop will focus on its warranties and shop rates, a landscaper on its service area and season, a physiotherapy clinic on its appointments and insurance receipts.
Keep anything you wouldn't put on your website out of the knowledge base: internal notes, margins, employee names, customer information, draft policies that haven't been adopted yet. Nothing sensitive either: health information, social insurance numbers, banking details. An assistant can repeat any passage it has been given. Treat every document you upload as public.
Generally, an AI assistant connected to your documents doesn't read them from start to finish. It looks for the passages that most closely match the question, then writes an answer based on those excerpts alone. Each passage therefore has to stand on its own.
Break your documents into short sections, each titled with the question it answers. "Can I return a used bike?" works better than "Special provisions." If a section answers three questions, turn it into three.
Write the complete answer, with its conditions and exceptions, in one place. Avoid:
Returns accepted under the usual conditions.
Instead, write:
You can return a new item within 14 days of purchase, with the receipt, if it hasn't been used. Clearance items and parts installed in our shop cannot be returned or exchanged.
The first version forces the AI to guess what the "usual conditions" are. The second gives it everything.
"See section 4.2," "as mentioned above," "check our Delivery page": these cross-references work on paper, not in an isolated excerpt. The AI may find the reference without finding the information. Instead, repeat the essentials where they're needed, and keep track of where those repetitions are so you can update them together.
Your customers don't talk about a "merchandise return policy." They ask whether they can "bring back" or "exchange" a helmet. Work those words into your sections. Be precise, too: "5 business days" rather than "a few days," "Monday to Friday, 8 a.m. to 5 p.m." rather than "business hours."
This is a common trap, and an easy one to avoid. An old PDF says delivery takes 10 days; this year's FAQ promises 5. Faced with both, the assistant may go with either one, and you won't know which. Before uploading anything, pick a single source for each topic and remove the others. When a policy changes, replace the document rather than adding a new version next to it.
Summer hours, holiday closure, a promotion that has ended, a revised rate: what was true in March no longer is in November. Put a review date in each document and set a calendar reminder, for example at every change of season. Group information that changes often into a short document that's easy to edit.
A warranty that has been photographed or scanned is an image, not text. Depending on the tool, it will be misread, partly read or not read at all. Even an ordinary PDF can cause trouble with columns or complex tables. Try this test: select the text in the PDF and paste it into a plain text editor. If nothing can be selected, or if the pasted text is unreadable or jumbled, retype the content into a simple text file.
Put the assistant to the test yourself before showing it to your customers.
A knowledge base is never finished. Three habits keep it useful:
Echo Support's assistant answers from your files, uploaded in the "Knowledge Base" screen (TXT or CSV, one at a time), and from the published articles in your help centre. If your Shopify store is connected, through the connector offered by invitation during its pilot phase, the store's policies and its ten most recently updated published pages are added too, resynced every day: the writing rules above apply to them as well. The assistant can then also search your published products and look up an order (the Shopify connector).
The assistant doesn't browse the rest of your website and doesn't draw its answers from your past conversations, which are only kept as long as your plan allows: since September 28, 2026, the older ones are deleted every night. Your documents and articles, on the other hand, are kept as long as the account exists.
TXT files are read as is, while CSV files are first converted to text by a language model. For important information, a well-structured TXT file or a help centre article remains the safest option. A file uploaded again under the same name replaces the previous one instead of sitting beside it. The number of files depends on your plan, and help centre articles have their own limit (the knowledge base documentation).
Articles are well suited to information that changes often: when you edit a published article, it's re-indexed, and an article moved back to draft or archived is removed from search. When a published article answers the question, the assistant can also offer the visitor a link to it.
Each search keeps at most five passages, then the assistant writes its answer with the instruction to stick to what it found. When nothing matches, it is instructed to say so and offer a human. If the visitor accepts, the conversation goes to your team and the assistant stops replying in it; the alert email only goes to members who have turned it on (handing over to a human). Some of this behaviour relies on instructions given to the model, not on guarantees: the tests above remain essential.
The app and its database are hosted in Quebec, in Beauharnois. Text processed by the AI goes to OpenRouter, in the United States, which routes it to DeepInfra (United States or Canada) or, failing that, Microsoft Azure (United States): your customers' messages, the excerpts of your documents the assistant looks up, and your CSV files when they're converted. For search, the text of your documents and articles also goes to Cohere, a Canadian company whose place of computation is not guaranteed to be Canada. Sensitive information is not masked before it is sent to the model: all the more reason to upload only what you would display publicly. Every subprocessor, what it receives and where it is.
The AI assistant and the knowledge base are included in the paid plans; the Free plan lets you try the chat widget, without AI.