v0.9.0

Changelog

Features 🏎️

• #640 - token_count() loads only the tokenizer, not the ONNX model, when lazy_load=True by @Mirochill @joein
• #643 - add lightonai/LateOn and the multilingual lightonai/mLateOn by @tekumara @joein
• #655 - add stemmer and stopwords parameters to Qdrant/bm25 for languages without a Snowball stemmer, such as Polish by @Mohith26 @joein
• #702 - add ibm-granite/granite-embedding-small-english-r2 by @Aryan-SINGH-GIT @joein
• #743 - add BAAI/bge-reranker-v2-m3 and its int8 version BAAI/bge-reranker-v2-m3-int8 by @joein
• #747 - speed up MUVERA document encoding for large k_sim values by @Pothan0 @joein
• #751 - add Qdrant/stella-en-400M-v5-doc-onnx for documents and the small query encoders Qdrant/constella-zero and Qdrant/constella-nano that search its embeddings by @joein @Dylancouzon
• #756 - add output_name to TextEmbedding.add_custom_model() to choose which ONNX output a custom model uses by @tuanzirwar @joein

Fixes 🔧

• #640 - fix BM42 and miniCOIL failing, and ColBERT keeping punctuation vectors, with lazy_load=True and parallel by @Mirochill @joein
• #642, #726, #727, #728, #729, #730, #732, #780, #787 - make model downloads more reliable: re-download corrupted cached files, retry on more network errors, time out on an unresponsive Hugging Face endpoint, and download built-in models from Hugging Face only (see Upgrade Notes) by @Dmheath1 @joein
• #670, #734 - fix a hang when a parallel run is stopped early, for example with break by @FU-max-boop @joein
• #721 - raise a ValueError for a batch_size below 1 by @adityaanikam @joein
• #723 - fix the licenses listed for Qdrant/colpali-v1.3-fp16 (gemma) and the Unicom models (mit) by @joein
• #725 - accept nearest-neighbor interpolation in Jina CLIP image preprocessing by @HuaTNA @joein
• #731 - match Jina's center crop in jinaai/jina-clip-v1 preprocessing, which was off by 1 px for some image sizes by @joein
• #733 - raise a ValueError when MUVERA encodes an empty multivector by @Pothan0 @joein
• #746 - fix center crop returning undersized images when it pads by an odd number of pixels by @HuaTNA @joein
• #754 - return float32 instead of float64 embeddings from mean-pooling text models, such as sentence-transformers/all-MiniLM-L6-v2 (regression in 0.6.1) by @joein
• #757 - fix loading models with .onnx_data weights from a huggingface_hub>=1.32 cache with onnxruntime>=1.24 by @joein
• #760, #761, #784 - fix Qdrant/colmodernvbert image embeddings in batches that mix aspect ratios, for portrait images, and for very thin images, and speed up its text queries (see Upgrade Notes) by @hulkbig @JesusMan0529 @joein
• #768 - return int64 indices from SparseEmbedding.from_dict({}) by @RAMZI0TO99 @joein
• #769 - fix a token ID overflow in Qdrant/bm25 query embeddings by @RAMZI0TO99 @joein
• #770 - respect cuda=False in parallel workers, which ran on the GPU when onnxruntime-gpu was installed (regression in 0.8.0) by @RAMZI0TO99 @joein
• #771 - respect do_resize=False in ConvNeXT image preprocessing by @RAMZI0TO99 @joein
• #773 - keep very small SPLADE weights by computing them with log1p by @RAMZI0TO99 @joein
• #775, #776 - honor the truncation and padding side from the model's tokenizer config by @hulkbig @joein

Upgrade Notes

• Built-in models now download from Hugging Face only, without the Google Cloud Storage (GCS) fallback some of them had. The GCS buckets stay open, so older versions keep working, and custom models still support both url and hf sources. GCS copies cached by older versions still load offline or when Hugging Face is unreachable, with a FutureWarning: load the model once online, then delete its fast-<model name> dir.
• We removed a stray 128-token limit from the Hugging Face configs of sentence-transformers/all-MiniLM-L6-v2 and Qdrant/bm42-all-minilm-l6-v2-attentions (now 256 tokens), and thenlper/gte-large and prithivida/Splade_PP_en_v1 (now 512). This applies to any fastembed version on a fresh download, and embeddings of texts over 128 tokens change. A cached model keeps the old limit until you delete it from the cache.
• Qdrant/colmodernvbert now loads model_v2.onnx and drops padding rows from its embeddings, so offline setups need to download it first. Re-embed images that were embedded in batches with images of other aspect ratios. The model needs onnxruntime>=1.19.
• LateInteractionMultimodalEmbedding.embed_image now defaults to batch_size=2 instead of 16, which also applies to Qdrant/colpali-v1.3-fp16.
• jinaai/jina-clip-v1 image embeddings change for some image sizes.

Thanks to everyone who contributed to this release @adityaanikam @Aryan-SINGH-GIT @Dmheath1 @Dylancouzon @FU-max-boop @HuaTNA @hulkbig @JesusMan0529 @Mirochill @Mohith26 @Pothan0 @RAMZI0TO99 @tekumara @tuanzirwar @joein