EvereplySign in →

Training your assistant

Your assistant only ever answers from the knowledge you give it. There are four ways to add knowledge, and you can mix all of them on the same assistant.

Source types

FAQ

Paste question-and-answer pairs. Each pair should be a separate block, one blank line apart, so it stays together as a single unit of knowledge:

Q: What are your opening hours?
A: Mon-Sat 9am-8pm.

Q: Do you deliver?
A: Yes, city-wide within 24h.

Text

Paste any document body — a policy, a product description, a procedure. Longer text is split into overlapping chunks automatically (see "How chunking works" below).

PDF

Upload a PDF up to 10MB. We extract the text and chunk it the same way as pasted text.

Website crawl

Give us a starting URL and we crawl up to 10 pages on that same site — the start page plus same-origin links we find on it — and turn each page's readable text into knowledge. Only public http/https pages are crawled.

How chunking works, in plain words

Behind the scenes, each source is broken into small pieces ("chunks") — an FAQ pair is one chunk, a long document is split into several overlapping chunks. When a customer asks a question, we search across all of an assistant's chunks for the most relevant ones and hand only those to the AI model to answer from. This keeps answers grounded in your actual content instead of the model "filling in" from general knowledge.

Tips for better answers

  • Write FAQs as clean Q/A blocks separated by a blank line — this is the single biggest quality lever.
  • Keep one topic per chunk; split a huge policy document into several shorter, focused sources.
  • Update a source and re-add it after you change a price or policy — the assistant only knows what's currently trained.
  • Test with real customer phrasing, not just the exact wording from your documents.

When the assistant doesn't know something

Rather than guess, the assistant says it isn't sure and asks for a name and contact so your team can follow up — this becomes a lead in your dashboard. This is intentional: an honest "I don't know" protects your customers and your reputation far more than a confident wrong answer.

Source statuses

  • processing — we're extracting, chunking, and embedding the source.
  • ready — the source is live and being used to answer questions.
  • error — something failed (e.g. an unreadable PDF or a crawl that couldn't reach the site); check the message next to the source and try again.
← OverviewInstalling the widget →