Making non-black-box embeddings: ask an LLM questions instead of using 1536-dim vectors

kieranklaassen · x · 2026-09-26

kieranklaassen shares a neat idea: build non-black-box embeddings by asking an LLM (via Jev) questions about each document and using the answers as the vector, instead of an opaque 1536-dimension embedding. For example, the email "I got charged twice this month, pls fix asap" becomes a readable vector like [iscustomer, urgent, aboutbilling, needsreply...]. He later corrects that a weighted dot product, not cosine similarity, should be used. The result is interpretable, tunable retrieval and classification.

Original post →

Making non-black-box embeddings: ask an LLM questions instead of using 1536-dim vectors

More from coding & agent