Turn text into vectors for semantic search, retrieval and clustering.
Quick picks
2 models
OpenAI's highest-quality embedding model. Returns 3072-dim vectors by default and supports reducing dimensions via the dimensions parameter. Outperforms text-embedding-3-small and the older ada-002 on MTEB and multilingual MIRACL retrieval benchmarks, for cases where accuracy matters more than cost.
embeddingretrievalrag
$0.156/1M in
OpenAI's small, low-cost embedding model. Returns 1536-dim vectors by default and supports shortening output dimensions via the dimensions parameter without retraining. Replaced text-embedding-ada-002 with better retrieval quality at a fraction of the price, and is the default choice for general-purpose semantic search and RAG.
embeddingretrievalrag
$0.024/1M in
Hugging Face
BAAI (Beijing Academy of AI) open-weight English embedding model with 335M parameters. Returns 1024-dim vectors and was a top MTEB English retrieval model on release. The v1.5 update improved similarity distribution so it works well without a query instruction prefix for symmetric tasks. A widely used open alternative to hosted embeddings.
embeddingretrievalrag
Currently not offered
Hugging Face
BAAI multilingual embedding model covering 100+ languages with an 8192-token context. M3 stands for its multi-functionality (dense, sparse and ColBERT-style multi-vector retrieval), multilinguality and multi-granularity over long documents. Returns 1024-dim dense vectors and is a strong open choice for cross-lingual and long-text retrieval.
embeddingretrievalrag
Currently not offered
Meta AI 650M-parameter protein language model trained on UniRef50 sequences. Feed it an amino-acid sequence and the per-residue hidden states act as learned protein embeddings, used for structure prediction, variant-effect and function tasks. This 33-layer checkpoint is the common balance of quality and cost in the ESM-2 family.
scienceembeddingresearch
Currently not offered
Hugging Face
Nomic AI open embedding model with a fully reproducible training pipeline (open weights, data and code). Supports an 8192-token context and Matryoshka representation learning, so you can truncate the 768-dim output down to 64 dims with graceful quality loss. Uses task prefixes like search_query and search_document.
embeddingretrievalrag
Currently not offered
Hugging Face
Sentence-transformers model fine-tuned from Microsoft PubMedBERT on PubMed title-abstract pairs by the NeuML team. Produces 768-dim sentence embeddings tuned for biomedical semantic search and similarity, and is the embedding backbone behind the paperai and txtai medical search tools.
scienceembeddingresearch
Currently not offered
Hugging Face
AllenAI document-level embedding model for scientific papers. Built on SciBERT and trained on the citation graph so that papers citing each other land close together. Feed it a title plus abstract and it returns one 768-dim vector per paper, useful for recommendation, clustering and citation-based retrieval.
scienceembeddingresearch
Currently not offered
Voyage AI
Voyage's general-purpose embedding model. 1024 dims, 32k context, strong retrieval performance.
embeddingretrievalgeneral
Currently not offered
Hugging Face
DMIS-Lab (Korea University) BERT-base initialized from English BERT and further pretrained on PubMed abstracts. Used as a feature extractor it yields 768-dim contextual embeddings tuned for biomedical text mining tasks such as NER, relation extraction and biomedical question answering.
scienceembeddingresearch
Currently not offered
Microsoft
Microsoft BiomedBERT (formerly PubMedBERT) pretrained from scratch on PubMed abstracts with a domain-specific vocabulary, rather than adapting a general model. As a feature extractor it gives 768-dim biomedical embeddings and set the original state of the art on the BLURB biomedical NLP benchmark.
scienceembeddingresearch
Currently not offered
Cohere's multilingual embedding model. Supports 100+ languages with separate search and classification modes.
cohereembeddingmultilingual
Currently not offered
Hugging Face
Alibaba (Tongyi Lab) general text embedding model. The v1.5 release extends the context to 8192 tokens and returns 1024-dim vectors, scoring competitively on MTEB while handling much longer inputs than typical 512-token encoders. A practical open model when documents exceed the usual short-context limit.
embeddingretrievalrag
Currently not offered
Jina's frontier multilingual embedding model. 570M params, 8192 ctx, 89 languages, Matryoshka dims 128-1024.
jinaembeddingmultilingual
Currently not offered
Hugging Face
Microsoft E5 multilingual embedding model with 560M parameters, initialized from XLM-RoBERTa-large and trained with weakly supervised contrastive learning. Covers around 100 languages and returns 1024-dim vectors. It expects query: and passage: prefixes on inputs and is a popular open model for multilingual semantic search.
embeddingretrievalrag
Currently not offered
Other
Mixedbread's open-source 335M embedding model. Top MTEB benchmark for English retrieval at release.
mixedbreadembeddingopen-weights
Currently not offered
Hugging Face
AllenAI BERT-base pretrained from scratch on 1.14M scientific papers (mostly biomedical and computer science) with its own scientific WordPiece vocabulary. Used as a feature extractor it gives 768-dim contextual embeddings tuned to scientific text, outperforming general BERT on tasks like NER and relation extraction in research corpora.
scienceembeddingresearch
Currently not offered
Voyage AI
Voyage's code-specialized embedding model. Up to 32k context, Matryoshka 256-2048 dims, int8/binary support.
embeddingcoderetrieval
Currently not offered
Their pages stay online, but they canβt be run at the moment.
Embedding models turn text β or sometimes images, code, or audio β into a fixed-length vector of floating-point numbers. Similar inputs land close together in the embedding space, dissimilar inputs land far apart. Reach for embeddings when building semantic search, retrieval-augmented generation (RAG), recommendations, or clustering.
Pricing is per-token, similar to text generation but typically 10-100Γ cheaper. On Railwail, OpenAI text-embedding-3-large costs $0.156 and text-embedding-3-small $0.024 per million tokens. Open-weights options (Jina V3, BGE, MxBai) cost only compute on your own infrastructure. A typical RAG corpus of 10 million tokens (around 20,000 documents) costs about $1.56 (large) or $0.24 (small) to embed once. Re-embedding on every model upgrade is the main long-tail cost.
The trade-off is dimension, recall, and price. Higher-dimensional embeddings (3,072 or 4,096 dims) capture more nuance but cost more to store and search. Lower-dimensional models (256-768 dims) cost ten times less and still recover the right document 90-95% of the time on most workloads. Use the high-dim flagship when retrieval quality is mission-critical (legal search, medical Q&A); use a budget model when you can tolerate the occasional missed result.
Watch out for chunk size: most embedding models perform best on chunks of 200-500 tokens. Embed an entire 50-page document as one vector and you lose the per-section meaning. Embed too small (under 50 tokens) and individual chunks become noisy. Pick a chunker that respects paragraph boundaries and adds a small overlap (10-20%) between chunks.
Watch out for multilingual mismatch: not every embedding model speaks every language equally. If your corpus is multilingual, pick a model whose training data covers your languages β Jina V3, Cohere Multilingual, and Voyage Multilingual are the safe defaults.
Top picks above cover the highest-recall flagship, the cheapest production model, the highest-dimensional option, and the fastest indexer.
Voyage 3 and OpenAI text-embedding-3-large currently lead on the MTEB benchmark for general English retrieval. Cohere Embed v3 multilingual leads for cross-lingual search. For code, Voyage Code 3 and CodeRankEmbed lead. Run a recall@k test on your own corpus before committing.
On Railwail, OpenAI text-embedding-3-small costs $0.024 and text-embedding-3-large $0.156 per million tokens. Open-weights models (BGE, Jina V3, MxBai) cost only compute when self-hosted. Embed the corpus once and amortize across millions of queries.
Most models optimize for chunks of 200-500 tokens. Use a chunker that respects paragraph boundaries, with 10-20% overlap between adjacent chunks. For very short queries (single sentences), some embedding models also support a 'query' vs 'document' mode that improves retrieval quality.
Standard sizes are 384, 768, 1,024, 1,536, and 3,072 dimensions. Higher dimensions capture more nuance but cost more to store and search. Many flagships now support Matryoshka representation β embed at 3,072 then truncate to any smaller size without re-embedding.
Multilingual flagships (Jina V3, Cohere Multilingual v3, Voyage Multilingual) cover 100+ languages with strong cross-lingual retrieval. English-only models lose 30-60% recall on non-English text. Always pick a multilingual model if your corpus or queries are not English-only.
Yes β dedicated code-embedding models (Voyage Code 3, CodeRankEmbed, Jina Code) outperform general models by 15-30% on code search benchmarks. They understand syntax, identifier naming, and cross-language similarity. Use them when building 'find similar code' or PR-review tooling.
Yes β CLIP and SigLIP variants embed images and text into a shared space, so you can search images by text query. Jina V3 also ships a multimodal variant. For pure image-image search, dedicated vision encoders like DINOv2 outperform CLIP.
Pgvector for Postgres-native workflows, Qdrant or Weaviate for high-scale standalone, Pinecone for managed simplicity, Milvus for very-large-scale on-prem. Railwail embeddings are compatible with all of them β pick the one that matches your existing infrastructure.
Every available model through one OpenAI-compatible API. Prepaid credits in USD, no subscription.
