LLMs as interpretable embeddings: named vectors and pairwise relation matrices
kieranklaassen · x · 2026-09-27
An interesting chain of embedding experiments:
- Named vectors instead of black-box vectors: kieranklaassen proposes skipping embedding models' 1,536 meaningless numbers and instead asking an LLM questions about each document; the answers form an interpretable vector (e.g. an email rated [iscustomer, urgent, aboutbilling, needsreply] → [1.0, 0.9, 1.0, 1.0]). Plain cosine similarity then surfaces the double-charge email for a "billing issues from customers" query, beating a newsletter.
- Pairwise relation matrix: trevin extends the idea, noting cosine similarity can't express relationships like "ticket A blocks ticket B" even if both are [bug, angry, churn risk]. He builds a matrix where rows/columns are items and each cell gets its own small vector from pair questions: same issue, A blocks B, B blocks A, duplicate. You read the cell itself — it's a score for one ordered pair, not for either item.
Key insight: classic vector similarity only encodes per-item attributes, not item-to-item relations; LLM-generated named vectors and pair vectors yield fully interpretable semantic search and relationship reasoning.
More from coding & agent
- Claude Code now wraps up cleanly when hitting the 5-hour usage limit — goyalshaliniuk · 2026-09-27
- Claude-generated Three.js 3D harbor scene wows devs with its detail — EricBuess · 2026-09-27
- Agent monitors flight cancellations for pennies using Mercator travel API in Grok Bot — jeff_weinstein · 2026-09-27
- Claude Opus 5.5 writes a 6:54 film as code, GitHub Actions renders the MP4 — EricBuess · 2026-09-27
- One prompt, 321 lines: Opus 5.5 builds a playable flight simulator in a single HTML file — EricBuess · 2026-09-27
- Claude Opus 5.5 one-shots a full retro Diablo-style RPG with graphics and audio, no generators — EricBuess · 2026-09-27