Kieran Klaassen uses Jev answers as embedding vectors
Four named dimensions instead of 1,536 anonymous ones, with one Jev call per question per document.
Kieran Klaassen is using Jev as an embedder. Instead of an embedding model returning 1,536 numbers that mean nothing, he asks Jev a fixed set of questions about each document and uses the answers as the vector. Then it is ordinary cosine similarity search over those vectors.
Every number has a name, so you can see why something matched. Need a new dimension? Add a question.
The worked example is an email in Cora. The text "I got charged twice this month, pls fix asap" scores the dimensions [is_customer, urgent, about_billing, needs_reply] as [1.0, 0.9, 1.0, 1.0]. A newsletter comes out at [0.0, 0.0, 0.0, 0.1]. A friend asking about lunch lands at [0.0, 0.1, 0.0, 0.7]. Query for "billing issues from customers" as the vector [1, 0.5, 1, 0.5] and, Klaassen says, the double charge comes out on top.
The point is that every number has a name. You can read the vector and see why a document matched, which you cannot do with a dense embedding. Need another axis, add another question. Want urgent items ranked first, change the query vector rather than retraining anything. Klaassen lists the same shape for other corpora he works with, [is_tutorial, about_ai, contrarian, beginner_friendly] for articles at Every, and [is_bug, angry, churn_risk, enterprise] for support tickets. He says he is trying it in Cora now to make search fast.
What it costs you
This is one person's design pattern posted the morning he thought of it, not a benchmark. There is no retrieval accuracy number here and no latency figure, so treat it as a technique to test rather than a result to cite.
For people building on it, the arithmetic at index time is the thing to look at before the schema. Each dimension is a question and each question is a Jev call per document, so a four dimension vector is four calls for every item you index, and widening the vector costs linearly. Diogo Almeida replied that this regime has worked for him in production, calling it tune-able RAG.
some of our biggest prod successes are in this regime - tune-able RAG
Know what Jev is good at before you bet on it
One email a day. Under three minutes. What builders found out about Jev yesterday.

