A search engine that refuses to stay on the server

What it does

Orama is an in-process search engine that indexes and queries data through full-text, vector, and hybrid modes from inside a JavaScript runtime. It recently added an AnswerSession class for RAG-style chat over indexed documents, turning the same library into a lightweight retrieval backend for generative UI. You declare a schema up front—vector dimensions included—and then insert, filter, and search through a straightforward API.

The interesting bit

Rather than bolting vector search onto a traditional text engine, Orama treats embeddings as a first-class column type and offers plugins that generate them at insert-time via TensorFlow.js or securely proxy OpenAI from the client. That means you can evolve from keyword search to a ChatGPT-like RAG experience without swapping out your index or adding a new service.

Key highlights

• Toggle between full-text, vector, and hybrid search with a single mode parameter on the search method
• Schema supports 10 data types, including geopoint and fixed-size vector[<size>] arrays
• Plugin ecosystem covers embeddings, secure client-side OpenAI proxy, analytics, and framework integrations for Vitepress, Docusaurus, Astro, and Nextra
• Built-in typo tolerance, exact match, field boosting, facets, and stemming across 30 languages
• Runs in browsers, servers, and edge workers under an Apache 2.0 license

Caveats

• Vector and hybrid search require supplying your own embeddings at query time unless you add the TensorFlow.js-backed embeddings plugin
• The RAG chat features rely on OpenAI APIs; the README recommends a secure-proxy plugin to handle those client-side
• Vector dimensions must be declared in the schema at initialization, so you need to know your embedding model’s size upfront

Verdict

A solid bet if you want modern search and retrieval inside a single deployable unit without managing a dedicated daemon. Look elsewhere if you need a distributed, multi-node search cluster or strictly offline LLM inference.