AI fundamentals

Embeddings are compression with opinions

An embedding turns messy meaning into numbers. That is useful, but it is not neutral, magic, or guaranteed to match how your users think.

~5 minAI
scroll
01 · the shape

Similarity is a product choice

Vector search feels objective because it returns scores. But those scores come from a model trained to notice some kinds of similarity more than others. Your corpus, chunking, and user language decide whether that is helpful.

search · weird neighbors

The useful signal is usually already there.

guardrails armed

Hit the button and watch the hidden AI surface area light up.

02 · the move

Look at the neighbors

Do not tune embeddings from aggregate scores alone. Inspect nearest neighbors for real queries, label misses, and decide which mistakes matter to the product.

evals/nearest-neighbors.js

A simple neighbor dump can explain a lot of strange answers.

03 · the check

Pair vectors with old-school filters

Metadata filters, keyword search, recency rules, and reranking can make vector search less mysterious. The best retrieval systems are usually hybrids.

AI launch switchboard
sloppy
shippable0
fragile: the model is still freelancing without enough checks

Flip the switches and watch the story turn into a tiny operating model.

0neighbors worth inspecting per query
0search modes: vector plus keyword
0human-labeled eval set
04 · the takeaway

Numbers still need judgment

Embeddings are powerful because they compress meaning. They are risky for the same reason. Keep the compression visible.

source trail