RAGLite is an open-source Python toolkit for building retrieval-augmented generation (RAG) systems on DuckDB or PostgreSQL, without heavyweight infrastructure. Developed by Superlinear, it has about 1,200 stars on GitHub, is licensed under MPL-2.0, and is actively maintained in 2026.
Most RAG stacks sprawl across separate vector databases, embedding services and orchestration frameworks. RAGLite takes the opposite path: a focused toolkit that stores documents, chunks, embeddings and vectors in a database you already run, DuckDB for local work or PostgreSQL (with pgvector, or SQLite) for deployment. It implements techniques that measurably improve retrieval quality, such as late chunking and late-interaction (ColBERT-style) reranking, plus a query adapter that rewrites user questions for better matches. Ingestion handles Markdown and PDF files, and built-in evaluation tools and an optional Chainlit chat interface make it easy to measure and demo a pipeline end to end.
RAGLite deliberately keeps the operational footprint small enough that a laptop prototype and a production service can share the same code path: swap DuckDB for PostgreSQL and the pipeline, chunking strategy and evaluation harness stay identical. It was created by Superlinear, a Belgian AI consultancy, and reflects retrieval patterns the team refined across client RAG deployments.
RAGLite is free and open source under the Mozilla Public License 2.0. You install it from PyPI and run it against your own database and LLM provider.
RAGLite is for Python developers who want a serious but lightweight RAG stack: strong retrieval techniques, a boring database underneath, and no separate vector infrastructure to operate.