Why Nearby Plots Share a View: Similarity and Distance Metrics

Opening Scene

Two plots sitting close together on a survey map tend to share something real — the same ridge line, the same river bend, the same morning light. That’s not a coincidence of the map; it’s the reason the map is useful at all. A surveyor doesn’t just record where a plot sits, but can measure precisely how far it is from any other plot, in a way that reliably predicts how much they’ll actually have in common.

Distance metrics give embeddings this exact same precise, predictive measurement.

In Plain English

A distance metric — cosine similarity, Euclidean distance, dot product, among others — takes two embedding vectors and produces a single number describing how close they are in coordinate space, a number that reliably correlates with how similar the underlying content genuinely is. Choosing the right metric for a given embedding model and use case is what makes “closeness” in that coordinate space actually mean something.

The Old Way

Before distance metrics over embeddings, measuring genuine similarity between two pieces of content relied on much blunter instruments:

Exact string matching could only say two things were identical or not, similar to a survey with no distance measurement at all, only a binary “same plot” or “different plot.”
Overlap-based measures, like counting shared keywords, treated proximity as a crude tally, similar to estimating distance by counting shared landmarks rather than actually measuring the ground between two points.
Manually defined similarity rules required a human to encode, in advance, exactly which features counted as “similar,” brittle and unable to generalize to content the rule’s author hadn’t anticipated.

This reliance on binary matches, crude overlap counts, and brittle manual rules is precisely what distance metrics over embeddings were built to replace.

What’s Changing (and Why AI Is the Reason)

Cosine similarity, the most common metric for text embeddings, measures the angle between two vectors rather than their raw distance, correctly capturing similarity even when vector magnitude varies for reasons unrelated to meaning. This directly extends the coordinate space introduced in Article 1 with the specific measurement that makes it useful.
Different distance metrics suit different embedding models and use cases, and modern embedding models are typically trained and evaluated against a specific metric, making metric choice a genuine correctness question rather than a free stylistic choice. Using the wrong metric for a given model can silently produce meaningless similarity scores, even though the underlying computation runs without error.
AI-assisted evaluation can now systematically test how well a given metric and model combination actually predicts genuine human judgments of similarity for a specific domain, replacing what used to be a largely intuition-driven choice with an empirically grounded one. This turns metric selection from guesswork into a measurable, improvable decision.

The Metaphor, Fully Extended

Land-Survey Element	Distance Metric Concept
The measured distance between two plots on the ground	The numeric output of a distance metric between two embeddings
A surveyor’s chosen measurement method — straight-line, road distance, or elevation-adjusted	A specific distance metric — cosine similarity, Euclidean distance, or dot product
Choosing the wrong measurement method for hilly terrain, producing misleading distances	Applying a distance metric mismatched to the embedding model, producing meaningless scores
Confirming that “close” distances actually predict shared views and features on the ground	Empirically validating that a metric’s similarity scores predict genuine content similarity
A trusted, standardized measurement method used consistently across an entire survey	A distance metric applied consistently across an entire vector search system

For Beginners: What to Actually Do

Learn your embedding model’s intended distance metric before assuming any metric will do — most models are trained with one specific metric in mind.
Get comfortable with cosine similarity as the default starting point for text embeddings, while staying open to other metrics for other content types.
Before trusting a similarity score, sanity-check it against a few pairs of content you can judge similarity for yourself.
Notice that a “bug” producing oddly unrelated search results is often a metric mismatch, not a problem with the embeddings themselves.

For Practitioners and Leaders: The Deeper Layer

Treat distance metric selection as tied directly to embedding model choice, not as an independent, interchangeable setting.
Use AI-assisted evaluation to empirically validate that a chosen metric genuinely predicts similarity judgments relevant to your specific domain.
Document the metric-model pairing your systems rely on, since this coupling is easy to break silently during a future model migration.
Recognize that getting this pairing wrong doesn’t throw an error — it just quietly degrades every search result built on top of it.

Quick Recap

Distance metrics turn embedding coordinates into a precise, predictive measurement of genuine content similarity.
Cosine similarity, Euclidean distance, and dot product are the most common metrics, each suited to different models and use cases.
Metric choice is tied to the specific embedding model in use, and mismatches silently degrade results rather than producing visible errors.
AI-assisted evaluation can empirically validate that a chosen metric genuinely predicts similarity for a given domain.

Where This Fits in the Series

Article 1 introduced embeddings as coordinates. This article covered the distance metrics that turn those coordinates into genuine similarity measurements. Article 3 looks at choosing the right instrument for the survey in the first place.