Every product in this category carries a list of technologies and most of the list is about delivery rather than about truth. Sorting them by whether they affect the one thing that kills a base, which is an answer that quietly stopped being correct, leaves four that matter and three that are worth having and not worth choosing on.
Matters: ranked retrieval
Matching a question to the page that answers it, rather than to pages containing the same words, is most of what a reader experiences. Whether it is done with a classic text index, embeddings or both is an implementation detail; whether the right page comes up in the first three results is the product.
Matters: per-article metadata
An owner, a review date, a state. Unglamorous, absent from a surprising number of products, and the difference between a base that ages gracefully and one that has to be audited by hand. If the fields do not exist, everything downstream becomes a convention somebody has to police.
Matters: version history with an approval
Being able to say what the article said on a given date, and who approved it saying that, is what makes a base usable in a regulated setting. Most products record edits; fewer record an approval alongside them, and that is the distinction to test rather than assume.
Matters: an API you would actually use
The base will be read by the support tool, the chat assistant and eventually something nobody has built yet. A read API with search, and a write API for the metadata, is what keeps the content from being trapped, and it is worth checking before rather than after two hundred articles exist.
Changes the demo: generated answers, translation, analytics
An assistant that composes an answer from your articles is genuinely pleasant and it inherits whatever staleness is underneath. Machine translation multiplies the review burden by the number of languages. Analytics dashboards measure attention rather than usefulness. All three are reasonable to want and none should decide a purchase.
Questions people ask about knowledge management technologies
Does the retrieval technique matter at all?
Only through its results. Load twenty of your own articles into two products, ask three colleagues to find five specific answers, and buy whichever they found them in. That test settles it in an afternoon and no architecture description can.
Is generated retrieval a reason to delay?
No. It changes how the answer is delivered and not whether the answer is true, and the ownership discipline it needs is the same discipline a base needed before it existed.
What about search across every system we own?
Useful late, when content genuinely lives in several systems that are not going to merge. As a first purchase it indexes an empty base perfectly, which is the same as nothing.
Which of the four is most often missing?
The approval alongside the version. Products aimed at internal wikis almost never have it, and companies in regulated work discover the gap at the point where somebody asks what the procedure said last March.