Plot twist: there are no transformer models, GPU clusters, or vector embeddings here.We built Glyph because we needed fast content intelligence for some of our own products.Then we made it fast.Glyph converts text into fixed-size signatures using MinHash and uses those signatures for similarity estimation, ranking, search, recommendations, guided token-chain autocompletions, and document matching. Everything runs in memory, with no embedding API or vector database required.In our rough benchmarks, Glyph s direct indexing implementation is approximately 87% faster than FuseJS for document-level fuzzy search.You can push performance further by using Glyph s higher-level Collections primitive and leaving LSH indexing enabled.What can you use it for?- Fast recommendations search: Keep a live in-memory index and retrieve candidates before hitting your database or disk.- Context retrieval for LLMs: Glyph Spotlight splits documents into chunks and ranks those chunks by Jaccard similarity, giving you a lightweight retrieval step before sending context to an LLM.- Preprocessing for vector databases: Deduplicate, filter, or cluster large corpora before spending money generating embeddings and storing vectors.The important distinction is that Glyph isn t trying to replace embeddings everywhere. If you need semantic understanding, embeddings are still useful.But a surprising number of retrieval problems don t need semantic embeddings in the first place. We built Glyph for those problems.Our first cloud products will be running Glyph under the hood. We ll share them here as they roll out.The library is open source and written in TypeScript: GitHub: https://github.com/Koda-OSS/Glyph Website: https://www.koda.sh/ Discord: https://discord.gg/Uc2Dnyb3EjI d especially like feedback from people who ve built search or recommendation systems and can poke holes in the approach.