LANDMARK · 7 MIN · UNDERSTAND
Embeddings
After this landmark, you can explain how semantic search finds documents that share no keywords with the query, and where that similarity misleads.
↗
This concept is shared, but the Everyday lens is less central here.
You can still explore it. We’re showing the shared explanation and a related practical view without hiding the knowledge.
An embedding is a list of numbers — often several hundred to a few thousand — that represents a piece of text as a point in space. The model producing them is trained so that texts with similar meanings land near each other. Once your documents are points, “find me things about parental leave” becomes a geometry problem: embed the question, find the nearest document points, return them. This is why semantic search retrieves a policy that only ever says “time off after the birth of a child”, with no keyword in common. The same trick works across modalities, which is how you search photos with a text description. The limitation to hold onto is that nearness is similarity of usage, not truth or relevance: a document arguing the exact opposite of your query sits very close to it, because opposites are discussed in nearly identical language.
This is the canonical concept. It stays the same across learner lenses so personalization never changes the underlying facts.
◌ EVERYDAY LENS
What this looks like for you
This is the machinery behind photo search that understands “dog on a beach”, and behind assistants that find the right note without the right word. It also explains why such a search sometimes returns something on-topic but useless — it matched the subject, not your intent.
MAKE A DECISION
Your semantic search over internal policies keeps returning a superseded 2023 version alongside the current one, ranked equally. What is the most likely cause?
CARRY THISSearch your organisation’s AI-powered search for a question whose answer you know. Note whether the top result answers it or is merely about the same topic.