EmbedAnything

EmbedAnything is an open source embedding pipeline written in Rust, maintained by Starlight Search. It handles ingestion, inference and indexing in one library: it reads text, PDFs, Markdown, images, audio, video frames and websites, splits them into chunks, generates embeddings and streams the resulting vectors directly into a vector database. Python and Rust interfaces are both documented. It is aimed at engineers building retrieval augmented generation systems and semantic search who want local embedding generation without a heavy PyTorch dependency, which keeps the memory footprint low and cloud deployment simple. It supports dense, sparse, ONNX, model2vec and late interaction embeddings. The Candle backend runs models including BERT, Jina, ColPali, Splade, ModernBERT, Qwen and rerankers, and the ONNX backend adds ColBERT, while cloud embedding providers OpenAI, Cohere and Gemini are also supported. Built in chunking covers semantic and late chunking, GPU acceleration is available, and vector database adapters let a pipeline switch its destination with minimal code. The project's distinctive feature is vector streaming: file processing, inference and indexing run on separate threads, so document preprocessing no longer waits on model inference and embeddings are written straight to the database rather than held in memory. The documentation includes benchmarks, adapters, notebooks and a public roadmap. EmbedAnything is free to use. The website publishes no paid plans or hosted tiers; it is distributed as an open source library through its GitHub repository and package documentation.