EmbedAnything

EmbedAnything is a local embedding pipeline for developers building semantic search and retrieval-augmented generation (RAG) applications. It turns documents and media into vectors for indexing in a vector database.

• Documents and media: Process PDFs, text, Markdown, websites, images and WAV audio, plus video through frame sampling.
• Embedding models: Choose dense, sparse or late-interaction embeddings with models such as BERT, Jina, Splade, ColBERT, ColPali, Model2Vec and Qwen3-Embedding.
• Text preparation and reranking: Use semantic or late chunking to prepare documents for retrieval, then rerank search results with models from Jina, BGE or Qwen.
• Vector database indexing: Stream embeddings directly to a vector database through adapters rather than collecting the full output in memory first.
• Application integration: Use the pipeline from Python or Rust, import files directly from AWS S3, or expose embedding generation through a self-hosted OpenAI-compatible API server.

The Apache 2.0 open-source project uses Candle and ONNX Runtime without a PyTorch dependency. It supports GPU acceleration and provides a Docker image for the server. Local models run on your hardware; optional OpenAI, Cohere and Gemini embedding services process requests in the cloud. Accessing private or gated Hugging Face models requires Hugging Face authentication.