Skip black-box embeddings: use LLM questions to build interpretable vectors
burny_tech · x · 2026-09-27
Kieran Klaassen shares a trick: instead of an embedding model spitting out 1,536 meaningless numbers, ask an LLM (Jev) questions about each document and use the answers as the vector — e.g. a double-billing complaint maps to [iscustomer, urgent, aboutbilling, needsreply] = [1.0, 0.9, 1.0, 1.0], then plain cosine similarity retrieves it. Delip Rao notes you can go further: treat the interpretable embedding as a target vector and learn a model to produce it, as in their 2023 stylometry paper — now easy for anything via Jev.
More from coding & agent
- Open-source GTA4 loading-screen template powers 'Grand Theft Alignment IV' AI-industry meme — jwt0625 · 2026-09-27
- The agents that actually stuck for me do one narrow thing, not everything — Cold_Hall_5384 · 2026-09-27
- AI now leads a quarter of tasks at Anthropic, up from 1% in February — victor_explore · 2026-09-27
- Non-engineer built an iPhone-only AI ops system, asks what breaks first — CellAgentLab · 2026-09-27
- $800 litter-picking robot MOSS hits V0.3, open-source V0.4 already printing — AIFlow_ML · 2026-09-27
- Matt Shumer buys a 3D printer for Opus 5.5, teases wild demos — mattshumer_ · 2026-09-27