StarlightSearch/EmbedAnything

Supports multimodal ingestion (PDFs, images, audio) with pluggable vector database adapters and multiple embedding backends including Candle, ONNX, and cloud models. Uses Rust's memory-safe concurrency and streaming architecture to separate document preprocessing from inference across threads via MPSC channels, eliminating sequential bottlenecks while preventing memory leaks. Offers dense, sparse, and late-interaction embedding strategies with built-in semantic chunking methods for RAG workflows.

1,174 stars. Actively maintained with 3 commits in the last 30 days.