mvlite - Fast Multi-vector Indexing and Retrieval

Build local or scalable RAG applications easily. A single durable object per user provides free partitioning for RAG.

Overview

mvlite is a python library focused on making multivector indexing, reranking, and retrieval fast on your local machine. For loads up to, say, a few million pages, this should be the fastest and easiest way to run ColPali/ColBERT/ColQwen/etc.

It's built atop uSearch.

Installation

To install mvlite, run:

pip install mvlite

Usage

from mvlite import MultiVectorIndex

index = MultiVectorIndex("demo.mvlite", ndims=3, nvectors=2)
index.add([[],[]])
index.add([[],[]])

index.search([[],[]])

Usage with ColBERT/ColPali