YC-backed Wondersearch claims OpenAI-level retrieval on FiQA with no embeddings, no vector DB
antoine_chaffin · x · 2026-10-28
Wondersearch (YC-backed) launched today with a provocative claim: "Embeddings were a mistake." The product lets companies search everything they've ever written and find the exact line.
The team says it matched OpenAI's best embedding model on the FiQA benchmark without using a single embedding, no vector database, no chunking, and no re-indexing. If it holds up, this is a direct challenge to the vector-retrieval default in today's RAG and enterprise search stacks — method details and independent replication will be worth watching.
More from Apps
- FreeSOV offers free AI share-of-voice tracking across 5 engines with your own API keys — gaganghotra_ · 2026-10-28
- Dev builds BasicMarks, an AI-tagged X Bookmarks manager with a knowledge graph view, seeking beta testers — bennash · 2026-10-28
- Convert non-narratable PDFs into audiobooks with ChatGPT — one user doubled productivity — Amaldevhari · 2026-10-28
- Anthropic is reworking Claude's plugins and connectors UI, soliciting user feedback — lydiahallie · 2026-10-28
- A peek at one user's Grok bot fleet: YouTube producer, medical bot, spam-call scrubber and more — petergyang · 2026-10-28
- Reddit user proposes a "Read Cursor" for ChatGPT so it knows where you stopped reading — doctordaedalus · 2026-10-28