The README quickstart used to read like this: generate 10,000 random 384D vectors, upsert them, query a random vector, look at the top-5 cosine scores. Top-5 typically came back at around 0.1, 0.09, 0.08. The numbers were correct. The conclusion was wrong: "this product returns garbage scores."
In high dimensions, random vectors have no neighbor structure. Cosine similarity converges to zero. Top-k of random against random is approximately the noise floor, by construction. The reader doesn't know that, and the product looked broken inside the first 3 minutes of trying it.
Two changes shipped together. First, replaced the synthetic vectors with 5,000 popular Wikipedia movie plots, sorted by Wikidata sitelinks for popularity, pre-embedded with sentence-transformers/all-MiniLM-L6-v2. Real text, real embeddings, real semantic structure. Second, bundled a quantized ONNX export of the same model in the SDK so the quickstart can encode arbitrary new query strings without a sentence-transformers install. samples.dataframe() returns the corpus, samples.encode(text) returns a query vector. The README's 6-line quickstart now ends with the entire Lord of the Rings trilogy as the top-3 hits for "a hobbit destroys a magic ring." Tested it on a few people who hadn't seen the product before -- they got it immediately. The quickstart was the demo all along, and I was demoing on noise.