Vector Databases

The Idea

A vector database stores embeddings and answers one question: which stored vectors sit nearest this one? Comparing against every row works until the collection grows, then an index groups neighbours so a query touches only a fraction.

Real-World Example

A textile warehouse shelved by shade and weave rather than by order number. A buyer holding a swatch walks straight to a neighbourhood of near-matches instead of unrolling every bolt in the building.

The Tradeoff

Approximate indexes swap recall for speed, and a missed neighbour never shows up as an error — the results simply look slightly worse. Exact search stays correct but scales linearly with the collection. For small corpora, brute force inside the database you already run is usually simpler and quick enough.

store = {"a": [1, 0], "b": [0, 1], "c": [0.9, 0.1]}
q = [1, 0]

def score(v):
  return sum(x * y for x, y in zip(q, v))

# rank the keys by similarity to q -> want ['a', 'c']
top = sorted(store, key=___, reverse=True)[:2]
print(top)