Skip to content

When sources contradict each other

Discover how glacom.AI handles information when two documents say different things and how to set priorities.

How it works

If you upload a PDF that says your product costs 10 euros, but your website says it costs 12 euros, the system must choose which data to give the customer. glacom® does not invent a middle ground, but follows a hierarchy of importance.

Priorities are managed like this: 1. Handwritten FAQ: they have absolute priority. If you write a specific answer in faq, it wins over everything. 2. Specificity: the most detailed information wins over generic information. 3. Recency: if importance is equal, the system tends to prefer the most updated data.

Tip

The best way to avoid errors is to delete old documents. Do not let outdated versions coexist with new ones.

What you need to do

  1. Clean your sources: remove files or web pages that contain data that is no longer valid.
  2. Use FAQ: if there is a critical point that is often mistaken, write a dedicated FAQ.
  3. Check content: follow the tips on how to write content to make information unique.
  4. Verify: use verification to test if the system responds with the correct data.

What glacom® does

The system analyzes sources through semantic search and applies priority filters to select the most reliable information before moving to the generation of the answer.

Limits

glacom® cannot independently know which document is "true" if both are written authoritatively and are up to date. If two different sources say two different things and neither is an FAQ, the system might alternate answers or show confusion.

Frequent problems

  • The bot says the old price? → Check if an old PDF or a hidden page on the site with that price still exists and remove it.
  • Can I prioritize a specific file? → Yes, by turning that information into an FAQ.
  • What happens if it doesn't find a unique answer? → The system might respond that it is not sure or follow the I don't know logic.

From the app

You can monitor wrong answers and report them to refine the training in training.