fastembed

Generates dense vector representations of text using the TextEmbedding class, which leverages ONNX Runtime for CPU-optimized inference instead of PyTorch. The library automatically downloads and caches pre-trained models (default: BAAI/bge-small-en-v1.5), applies tokenization and pooling strategies (mean, cls, last-token), and supports batch processing with data parallelism for efficient multi-document embedding at scale.

Generates sparse vector representations using the SparseTextEmbedding class, supporting multiple sparse embedding strategies (SPLADE, BM25, BM42) that produce high-dimensional vectors with mostly zero values. These sparse embeddings are designed to integrate with traditional keyword-based search systems, enabling hybrid search by combining dense semantic vectors with sparse lexical matching in a single retrieval pipeline.

Designed with minimal external dependencies (primarily ONNX Runtime and numpy), avoiding heavy frameworks like PyTorch or TensorFlow. This lightweight design enables deployment in resource-constrained environments such as AWS Lambda, Google Cloud Functions, and edge devices where package size and memory limits are strict. The library's total package size is <50MB, compared to 500MB+ for PyTorch-based alternatives.

Generates token-level embeddings using the LateInteractionTextEmbedding class, which implements the ColBERT architecture to produce embeddings for each token in a document rather than a single aggregate embedding. This enables fine-grained matching where query tokens are compared against all document tokens, allowing relevance scoring based on the best token-pair matches rather than document-level similarity.

Generates dense vector representations of images using the ImageEmbedding class, which leverages CLIP and similar vision-language models via ONNX Runtime. The class handles image loading, preprocessing (resizing, normalization), and batch inference to produce embeddings that capture visual semantics in a shared embedding space with text embeddings, enabling cross-modal search.

Generates token-level embeddings for document images using the LateInteractionMultimodalEmbedding class, implementing the ColPali architecture to produce per-patch embeddings from document images (PDFs, scans). This enables fine-grained matching where query tokens are compared against visual patches in documents, supporting retrieval of specific content within document images without OCR.

Scores pairs of texts (query-document, question-answer) using the TextCrossEncoder class, which applies transformer models that jointly encode both texts to produce relevance scores. Unlike bi-encoders that embed texts independently, cross-encoders directly model the relationship between text pairs, enabling accurate reranking of retrieval results or scoring of candidate answers without embedding the entire candidate set.

Automatically downloads pre-trained embedding models from Hugging Face Model Hub and caches them locally using a configurable cache directory. The system handles model versioning, integrity checking, and lazy loading, allowing developers to specify models by name (e.g., 'BAAI/bge-small-en-v1.5') without manual download management. Cache location defaults to ~/.cache/fastembed but is configurable for containerized or restricted-filesystem environments.

Processes large batches of documents efficiently using data parallelism, where the library automatically splits input batches across available CPU cores or GPU devices. The implementation uses ONNX Runtime's built-in parallelism and optional multi-threading to maximize throughput, allowing developers to embed thousands of documents with a single function call while the library handles batching, device allocation, and result aggregation.

Provides optional GPU acceleration through a separate fastembed-gpu package that replaces CPU ONNX Runtime with CUDA-optimized inference. When installed, the library automatically detects available GPUs and routes inference to GPU devices, providing 5-10x speedup for embedding generation. The GPU implementation maintains API compatibility with CPU version, requiring only package installation change without code modifications.

Provides a unified Python interface supporting 50+ pre-trained embedding models across multiple architectures (dense, sparse, late-interaction, multimodal) without requiring model-specific code. The library abstracts model differences through consistent class APIs (TextEmbedding, ImageEmbedding, etc.), allowing developers to swap models by changing a single parameter while maintaining identical inference code. Supported models include BAAI BGE, Sentence Transformers, SPLADE, ColBERT, CLIP, and ColPali variants.