How training and retraining work

From source to answer

Everything your bot knows starts as a source: a website page, a sitemap, an uploaded file, an OKF bundle, or a text snippet. Training turns those sources into something the bot can search.

• Sources. The items you add on the Website Links, Files & Data Sources, and Text Snippets tabs.
• Pages. Content is cleaned and counted in pages. One page is 2,500 cleaned characters, so a long document counts as several pages.
• Chunks. Each page is split into smaller passages that can be matched against a question.

When a visitor asks something, Matram retrieves the chunks that best match the question and passes them to the AI, which writes the answer from those passages. Follow-up questions are rewritten using the recent conversation, so "how much does that cost" still finds the right passages.

Answers cite their sources

Answers are grounded in your trained content, and the widget shows up to 4 source pills under an answer so visitors can see where it came from. You can hide the pills if you prefer.

In the default strict answer style, the bot declines questions your content does not cover instead of guessing. See the Response Engine and answer styles for how to change that behaviour.

Page allowances by plan

Plans include different amounts of trained content: the Basic plan includes 500 pages, Standard includes 2,500, and Pro includes 10,000. These are the advertised tiers; the one hard stop in Matram is the monthly message quota, described in plans and usage limits.

Automatic resync

Sources resync automatically so answers keep up with your content. The cadence depends on your plan:

• Basic: monthly.
• Standard: weekly.
• Pro: daily.

Every source also has a manual Resync button, so you can refresh it the moment something important changes. Text snippets are the exception: they never auto-resync and only change when you edit them.

Removing a source

Removing a source removes its content from the bot. Once it is gone, the bot no longer retrieves or answers from it. Use this when a document is outdated or a page should never have been trained.