The semantic space

By now you have seen points and neighbors and arrows. Zoom out: the embedding space as a whole is a geography of meaning. Regions are topics. Distances are relatedness. Directions are relationships. Everything a corpus knows about how words relate, drawn as one enormous map.

This reframe matters because modern AI systems navigate this map as their core operation. Semantic search does not match keywords; it embeds your query and finds the nearest documents. Recommendation engines find items near your taste. RAG systems (a later protocol) retrieve knowledge by distance in exactly this kind of space.

In a large share of production AI, the map is the product. The model generates the coordinates; the business logic is "measure distances, return the closest".

The standard ruler in these spaces is cosine similarity: the angle between two vectors, ignoring their lengths. Two vectors pointing the same way score 1, unrelated directions score near 0. Length tends to track word frequency more than meaning, so the angle is the cleaner signal.

You do not need the formula in your pocket; you need the operational meaning:

• cosine ≈ 0.9: same neighborhood, near-synonyms or tight topic partners.
• cosine ≈ 0.5: same city, related domain, would co-occur in an article.
• cosine ≈ 0.0: different planets.

One practical warning from real systems: "close" means RELATED, not SAME, and not even compatible. "love" and "hate" sit close (both emotions, shared contexts) while being opposites. A search system that treats proximity as agreement will cheerfully retrieve the exact opposite of what the user meant. The map tells you the district; it does not read the street signs for you.