Vectra
A provider-agnostic Node.js SDK for retrieval-augmented generation that has grown a second half: a context and memory layer that decides what a model sees on a given turn under a fixed token budget.
Vectra is a Node.js SDK for end-to-end retrieval-augmented-generation pipelines, covering embedding, chunking, storage, retrieval, reranking and generation, with every stage explicit and swappable rather than defaulting silently. It supports several embedding and generation providers, seven vector store backends, and retrieval strategies including hypothetical document embedding, multi-query expansion, hybrid reciprocal-rank fusion, and maximal marginal relevance.
A separate context-and-memory layer assembles what a model should see on a given turn under an explicit token budget through one call, and keeps a bi-temporal fact store where a contradicted fact is marked invalid rather than deleted. It ships a CLI with the same capabilities as the SDK, plus a local web UI for configuration and observability, and is licensed under MIT.