Non-black-box embeddings: use an LLM's answers to named questions as your search vectors
kieranklaassen · x · 2026-09-26
A practical trick from Every's Kieran Klaassen: instead of embedding models spitting out 1,536 meaningless numbers, ask an LLM (Jev) a fixed set of questions about each document and use the answers as the vector.
- An email "charged twice, pls fix" becomes [iscustomer, urgent, aboutbilling, needsreply] = [1.0, 0.9, 1.0, 1.0]
- Then plain cosine similarity works: query "billing issues from customers" as [1, 0.5, 1, 0.5] surfaces the double-charge email
- Every dimension has a name, so matches are explainable; add a question for a new dimension, tweak the query vector to re-rank
- He's applying it to Cora's email search, Every's articles, and support tickets ([isbug, angry, churnrisk, enterprise])
More from coding & agent
- Datasette 1.0a40 patches permissions bypass that leaked private rows via trailing newline — JeremyCMorgan · 2026-09-26
- Hamel Husain: make AI outputs easier to evaluate by redesigning the product, not the evals — HamelHusain · 2026-09-26
- Exa launches Agent Ultra, an agent-swarm deep research tool claiming SOTA — yoimnotkesku · 2026-09-26
- YC Backs 'Spend More Compute on Validation': Greptile Launches Plus and Apex Review Tiers — ycombinator · 2026-09-26
- OpenAI Showcases Proaction Using GPT-6 Astra to Build Fleet-Management Agents Faster — OpenAIDevs · 2026-09-26
- Prompt-to-Promo Video: Deterministic Frame Rendering Is the Real Unlock — ezshine · 2026-09-26