Glyph is a lightweight, open-source content intelligence toolkit for fast fuzzy search and similarity matching. It turns text and other content into compact fingerprints that can be indexed, compared, grouped, and queried without requiring vector databases or large models.
That sentence is the product. The interesting part is why it exists.
The bet
Most “smart search” stacks assume you need embeddings. You stand up a model, you store fat vectors, you pay for similarity at query time. That is a fine answer if your problem is open-ended language. It is a lot of machinery if your problem is this document is kind of like that one.
Glyph starts from a smaller claim: MinHash is enough. Tokenize, fingerprint, compare overlap. Jaccard-ish similarity in a tiny signature. No GPU. No embedding API. No second database if you do not want one.
The API stays small on purpose: create a glyph, compare two glyphs, put glyphs in an index, search. Around that core sit a few higher-level primitives; Spotlight for “which part of this document actually matched,” Collections for combining fields, and the obvious extensions (recommendations, fuzzy query). The goal is a composable toolkit, not a platform you have to adopt wholesale.
What we were optimizing for
Precision was never the north star. Work per query was.
A fingerprint is a lossy sketch. Two similar sentences should collide in enough slots to look related. Two unrelated essays should not. You give up some nuance; sarcasm, rare proper nouns, the one sentence that would have saved the match in exchange for something you can compare in microseconds and serialize to a string.
Spotlight came from the same instinct. A whole-document glyph can smear a long essay into mush. Chunking, fingerprinting the pieces, and ranking those pieces is how you keep MinHash honest without pretending it is a language model.
Serializing glyphs was never a research problem. It was a shipping problem. If the fingerprint cannot live in query.json, a cache, or a column, it is a toy.
It is running here
This portfolio uses Glyph for search and related documents. The fingerprints are generated at build time; the server ranks at request time. Glyph never runs in the browser.
Correction: Glyph never runs in the browser for search, Glyph spotlight is used by live captions.
For the technical walkthrough, the Koda-OSS Glyph-Blog template this site is forked from is the better map. The repo, Medium, and Product Hunt cover APIs, knobs, and benchmarks. This series is about the why.
Next: how we tricked ourselves into better fingerprints without teaching every developer MinHash.