AI processes text much like a person does: it gets confused by vague wording, loses context when sources contradict each other, and can’t guess what isn’t there. The main principle is simple:
What Gramax AI Search Takes Into Account
Ignores | Takes into account | Why |
|---|---|---|
Links between articles | Gramax splits all text into small chunks and looks for relevant information chunk by chunk. You don’t have to connect pieces of text explicitly — search will pick the right answer from the text itself | |
Text like: “See below” | Hierarchy:
| If the information “below” sits under a logical heading, search will find it without any extra pointers |
| Text next to an image | AI search can’t read text from images. If information from a picture should show up in search results, add it to the text or to the caption. |
Old versions of an article | The current version | Search returns results from a specific branch and the latest version of the article. This keeps outdated versions and drafts out of the answers. |
Text like: “For HR only” | Access permissions | If you’ve set up access permissions for catalogs, search only shows what a particular user is allowed to see. |
Coming soon! |
Tips
Logical Hierarchy
When splitting content into articles, make sure no article ends up in a section it has nothing to do with. The hierarchy is easy to test on your colleagues: ask them to find a specific article using only the section tree.
Informative Headings
Headings help AI build a hierarchy of your content: it figures out what the text under a heading is about and uses the heading as a summary.
Write headings from the user’s point of view. A lawyer thinks in terms of a “benefits package,” but an employee will search for “What benefits do I get when I have a baby?”
Bad | Good |
|---|---|
Additional | Parental benefits |
Information | Housing allowance for relocated employees |
A Single Source of Truth
If the same rule is described slightly differently in several articles, search may return contradictory answers. If you really need the same text in several places:
- Move it into a separate article and link to it where needed.
- Create a fragment and add it to every article.
Consistent Terminology
If your articles use different terms for the same thing, search will most likely still figure it out and return relevant answers. But if, say, five articles say “Support” and one says “Assistance programs,” context can get lost.
Up-to-Date, Consistent Content
Search can’t make up for missing knowledge: if the information isn’t in the knowledge base, the model can’t guess it, and the answer will be limited. That’s why it’s important to keep adding to your knowledge base and review it regularly: update outdated rules, record process changes, and remove contradictions between sections.
Why Use RAG When Your Content Is Already Good?
A good structure helps people find an article when they roughly know what they’re looking for. RAG solves a different problem: it answers a question even when a person doesn’t know which article to open, or doesn’t want to read at all.
For example:
- “I’m going on a 5-day business trip. What’s my per diem, and do I need approval?” A person would have to go through several articles and piece the answer together. AI search does it for them.
- A new employee doesn’t know the company’s jargon and won’t guess that relocation info lives in the “Workforce Mobility” section. AI search understands context and can search by synonyms.
- A manager doesn’t want to read — they want a quick answer.
A good knowledge base and AI search don’t compete: structure helps people who read, and at the same time makes search more accurate. A poorly organized knowledge base is a problem for both.