The open-source answer engine that packs full-text, vector, and hybrid search into a library lighter than the button you clicked to get here.
Here is a fact about the search-infrastructure business that sounds like a joke but is not: for a long time, if you wanted to add a good search box to your website, the first thing you had to do was stand up a cluster. Buy servers, or rent them, or sign a contract with a company that had done both and would now charge you per query, per record, per month, in a pricing table with an asterisk. Search was heavy. Search was a project.
Orama's opening argument is that this is silly. The core of Orama - the actual search engine, plus the pipeline that lets a large language model answer questions from your data - ships in under two kilobytes and runs, if you want it to, entirely inside a web browser. No cluster. No network call to a vendor. The database is the tab you already have open.
This is the kind of claim that invites eye-rolling, because software people say "lightweight" the way restaurants say "artisanal." But the tell here is who is running it. The official Node.js website uses Orama for search. So does tanstack.com, and solidjs.org, and jsr.io - the JavaScript registry. These are not toy sites. They are documentation platforms serving millions of queries a day to some of the most demanding users on earth, which is to say, developers, who complain loudly and in public when search does not work.
The company that makes this is OramaSearch Inc., incorporated in 2023 by Issac Roth and Michele Riva. It did not start life as a company. It started as an open-source project called Lyra, the sort of thing an engineer builds because the existing options annoy them. Lyra got stars on GitHub - the internet's applause meter for code - and then it got a lot of them, and at some point the reasonable thing to do with a project that thousands of people depend on is to put a company around it so it does not disappear.
So they renamed it Orama, put Roth in the CEO chair in San Francisco and Riva - a conference speaker, a book author, a former contributor to the committee that standardizes JavaScript itself - in the CTO chair in Italy, and set about the genuinely hard problem, which was never the search box. The hard problem was everything that has happened to search since 2023.
Figures compiled from public GitHub data and company materials; latency and size figures are as reported by Orama and are approximate.
The core engine. Full-text, vector, and hybrid search plus a RAG pipeline, in TypeScript, with no external dependencies. Drop it in a browser, a server, or an edge function.
A Rust-based runtime that bundles the search engine, a vector database, embedding generation, and an LLM interface into one deployable unit. Build an answer engine or a copilot without wiring five services together.
The hosted version: hybrid search, automatic embeddings, analytics, and webhook APIs, sold on a flat fee for unlimited queries rather than per-query metering.
Grounded, cited AI answers pulled from your own content - with confidence thresholds and guardrails aimed at keeping the model from inventing things.
Everyone is bolting a language model onto their product this year, which means everyone is discovering the same unglamorous truth: the model is only as good as the paragraph you hand it. Retrieval - finding the right chunk of text, fast - is the part that decides whether an AI answer is useful or embarrassing. Orama built that part first and the model second.
With OramaCore, the company says it pushed time-to-first-token - the delay before an AI answer starts appearing - from roughly five seconds down to about one. Five seconds is long enough to make a user leave. One second feels like the machine already knew. Nobody writes a thank-you note for fast retrieval. They just stay, which is the entire point.
The most convincing endorsement in developer tooling is not a testimonial - it is a dependency. When an open-source project as scrutinized as Node.js puts your search on its official docs, it means people who read source code for sport looked at yours and shipped it. Orama's adopter list reads like a tour of the modern JavaScript ecosystem.
Based in San Francisco. A repeat entrepreneur with a long run in technology and venture-backed companies, now running the commercial side of Orama.
Based in Italy. Conference speaker, author of a book on modern web apps, and a former contributor to ECMA International, the body that standardizes JavaScript. He is the one who started what became Orama.
Riva announces the company - the commercial home of the open-source Lyra project, rebranded Orama.
The company raises a reported seed round and builds out its cloud search product. (Figure approximate.)
A from-scratch runtime combining a search engine, vector database, embeddings, and an LLM - positioned as part of the Orama 4.0 platform.
The strange and interesting decision at the center of this company is that it gives away the crown jewels. The search engine, the vector database, the runtime - open source. The usual instinct in infrastructure is to hide the good code and rent access to it. Orama's bet is the opposite: that adoption and trust compound faster than any moat you build by being secretive, and that a developer who can read every line of your engine is a developer who will actually deploy it.
The monetization sits one layer up, in Orama Cloud, sold on a flat fee for unlimited queries - a quiet dig at the per-query meters that make search bills unpredictable. It is a small company, around fourteen people, remote-first, building in public. The ambition is to be the boring, invisible layer underneath a lot of other software. In infrastructure, boring is the highest compliment there is.