Meaning Map
Pick three words and watch the map do the arithmetic that every AI search engine runs silently: A − B + C. On a set of precomputed word vectors, king − man + woman lands exactly on queen, and paris − france + italy lands on rome — not by magic, but because each word is a point in space and meaning is a direction you can subtract. You see the target coordinate, the straight-line distance to every other word, and the nearest neighbour that answers the analogy. Runs entirely in your browser (0 uploads, works offline).
How to use it
- Load a preset — king − man + woman or paris − france + italy — or pick your own A, B, C from the dropdowns.
- Read the equation bar: it shows
A − B + C =the nearest word the math produced. - Look at the map: your three input words, the amber target marker where the arithmetic landed, and the dashed line to its nearest neighbour.
- Scan the ranked list to see the top four neighbours and their exact Euclidean distances, then open the X-ray to see why the offset works.
What this clears up (the fundamentals)
- A word is a coordinate, not a string — an embedding turns each word into a point in vector space, so "similar" becomes a measurable distance instead of a vibe.
- Meaning is a direction you can do algebra on — the offset king − man is roughly the "royalty without the gender" vector; add woman and you arrive at queen. Same trick moves a country to its capital.
- "Semantic search" is just nearest-neighbour — find the closest points to a query vector. That is the whole engine behind retrieval, RAG, and recommendation, minus the marketing.
- Distance is the score — the ranked list is sorted by Euclidean distance; the smallest number is the best match. No model is deciding — geometry is.
Where it's used
This is a miatz build-lab concept playable: play the analogy here, then learn to build the embedding lookup, the vector-arithmetic step, and the nearest-neighbour ranking yourself. It sits inside AIE-202's embeddings & vector stores sub-module as the fastest way to feel why retrieval systems work before you wire one up.
FAQ
Why does king − man + woman equal queen?
Because the vectors are placed so that king = royalty + male and queen = royalty + female. Subtracting man removes the "male" part and adding woman adds the "female" part, leaving royalty + female — which is exactly queen's coordinate.
What is the distance number in the ranked list?
The straight-line (Euclidean) distance in the 3D space between the target coordinate and each word. Smaller means closer in meaning; the top of the list — distance 0 for the exact analogies — is the answer.
Are these real word embeddings?
No — they are a small set of hand-chosen vectors designed to make the classic analogies land cleanly and stay readable. Real embeddings have hundreds of dimensions and are learned from text, but the arithmetic and nearest-neighbour idea are identical.
Why are only three words allowed in the analogy?
Analogy arithmetic is defined as A − B + C: take a relationship (A minus B) and apply it to a new word (C). Three inputs are all the "king is to man as ? is to woman" pattern needs.
Is anything uploaded?
No. Every step — the subtraction, the distances, the ranking — happens in your browser against baked-in vectors. Nothing is transmitted, stored, or logged. Turn off your Wi-Fi and it still works.
Limits
This runs on precomputed, illustrative vectors — about eighteen words in three dimensions, hand-placed so the textbook analogies resolve exactly. It is a teaching model, not a live embedding model: you cannot type arbitrary words, dimensions are collapsed to make the picture legible, and real-world embeddings encode far messier relationships. The concept it makes visible — meaning as coordinates, similarity as distance — is exactly what production retrieval systems rely on.
Related
Part of the miatz playgrounds. Explore the rest from the playground catalog.
Bookmark this page (Ctrl+D, or ⌘D on Mac) — it works offline the next time you need it.