EmbedAnything
A minimalist, multisource, multimodal local embedding pipeline built in Rust that generates embeddings from text, images, audio, and PDFs and streams them into a vector database with low memory use.
EmbedAnything is a modular, local embedding pipeline built in Rust for text, images, audio, PDFs, websites, and other media, streamlining the process of generating embeddings from varied sources and streaming them into a vector database. It supports dense, sparse, ONNX, model2vec, and late-interaction embeddings, and has no dependency on PyTorch, which keeps its memory footprint low and its cloud deployment simple.
Vector streaming separates file processing, indexing, and inference onto different threads using Rust's MPSC channels, so embeddings are saved directly to the vector database rather than held in memory, reducing pipeline latency and improving throughput. Backends include Candle, ONNX, and cloud models, with GPU acceleration and built-in chunking methods such as semantic and late chunking.
A prebuilt Docker image is published as starlightsearch/embedanything-server, and the project has Python and Rust API documentation and example notebooks.