Vector search answers a different question from keyword search. Instead of asking “which documents contain these words?”, it asks “which document vectors are closest to the query vector?”
This playground makes that retrieval loop visible. It uses a tiny, hand-authored knowledge base and a limited browser-only “embedder” so you can inspect every step. The map’s horizontal and vertical axes are understandable teaching dimensions; a real embedding space usually has hundreds or thousands of learned dimensions that cannot be drawn directly.
The toy embedder recognizes only the concepts used by the example queries. Real systems use a trained embedding model, but the retrieval flow is the same: embed the query, compare it with stored vectors, apply filters, sort, and return the top results.
Try this:
- Choose a paraphrased question such as “Why was I billed twice?” and notice that it finds an article titled “Resolve a duplicate card charge” without requiring the same wording.
- Change top-k to control how many nearest documents survive the ranking step.
- Apply a metadata filter and watch documents become ineligible before ranking.
- Switch the similarity metric and watch the ranking change. Euclidean distance follows the line between two points. Cosine similarity draws both vectors from the origin so you can see the angle and its cosine value.
- Select any result to inspect its vectors and the exact score calculation.