🤖 Model
replicate/all-mpnet-base-v2
Convert text into 768-dimensional embeddings for semantic search, retrieval, clustering, and similarity matching. Accept...
← All guides
Text Embedding API
A text embedding API turns text into a vector so you can compare meanings numerically — the foundation of semantic search, RAG retrieval, deduplication, clustering, and recommendations.
On Replicate the practical lineup: all-mpnet-base-v2 (2.4M+ runs, the dependable English default), BGE-Large (the stronger English retriever), BGE-M3 (multilingual + multi-mode), and CLIP features (163M+ runs — embeddings that bridge text and images).
Quick Recommendation
For English semantic search or a first RAG pipeline, start with replicate/all-mpnet-base-v2 — fast (~1s example runs), well understood, supports batch input via text_batch.
Step up to nateraw/bge-large-en-v1.5 when retrieval quality matters — BGE models consistently outrank MPNet on retrieval benchmarks, and example runs are sub-second.
Use lucataco/bge-m3 when your corpus is multilingual (100+ languages) or you want its unusual trick: dense, sparse, and multi-vector retrieval modes from one model (embedding_type input).
Use andreasjansson/clip-features when the job involves images — CLIP embeds text and images into the same space, so "find images matching this sentence" becomes a vector lookup.
Keyword-Matched Use Cases
| Search intent | What the user likely needs | Best starting point |
|---|---|---|
| Text embedding API | General-purpose sentence vectors | all-mpnet-base-v2 |
| Embeddings for RAG | Best retrieval quality for chunks | BGE-Large |
| Multilingual semantic search | One model across languages | BGE-M3 |
| Hybrid search (dense + sparse) | Both vector and lexical-style signals | BGE-M3 embedding_type |
| Image search by text | Text ↔ image similarity | CLIP features |
| Duplicate / similarity detection | Cheap fast vectors at scale | all-mpnet-base-v2 or BGE-Large |
Model Comparison
| Model | Runs on Replicate | Example speed | Languages | Distinctive strength |
|---|---|---|---|---|
| replicate/all-mpnet-base-v2 | 2.4M+ | ~1s | English | Dependable default, batch input |
| nateraw/bge-large-en-v1.5 | 312K+ | <1s | English | Stronger retrieval quality |
| lucataco/bge-m3 | new on Replicate | slower per call | 100+ | Dense + sparse + multi-vector in one model |
| andreasjansson/clip-features | 163M+ | <1s | English | Shared text–image embedding space |
Choosing in practice
RAG pipelines: embedding quality caps retrieval quality. BGE-Large over MPNet is one of the cheapest upgrades available. Keep chunks within each model's context limit and embed queries with the same model as documents — mixing models breaks similarity.
Multilingual corpora: don't machine-translate then embed — BGE-M3 embeds 100+ languages into one space, so a German query can retrieve French documents directly. Its max_length also stretches to long passages, reducing chunking pressure.
Hybrid retrieval: BGE-M3's sparse mode approximates keyword-style matching from the same model that produces dense vectors — a pragmatic path to hybrid search without running a separate BM25 stack.
Text-image: CLIP's 163M+ runs reflect how central it is — product search, moderation pipelines, and image dedup all reduce to CLIP vectors plus a vector database.
Note the Replicate wrapper detail: lucataco/bge-m3 takes two sentence lists (sentences_1, sentences_2) and is oriented to comparing them; for bulk index-building, batch your calls accordingly.
🤖 Model
nateraw/bge-large-en-v1.5
Convert English text into dense vector embeddings. Accepts lists of text strings and returns normalized embeddings for s...
🤖 Model
BGE-M3 — Multilingual Embedding Model for Search and Retrieval
BGE-M3 is a multilingual embedding model that supports over 100 languages. It generates dense vector representations fro...
🤖 Model 🖼️
andreasjansson/clip-features
Generate CLIP ViT-L/14 embeddings from text and images for cross-modal similarity search and retrieval. Accept text stri...
Practical Decision
Default to BGE-Large for English retrieval and RAG; keep all-mpnet-base-v2 where ecosystem familiarity or batch shape fits better. Go BGE-M3 for multilingual or hybrid retrieval, and CLIP features whenever images enter the picture. Above all: embed queries and documents with the same model, always.
FAQ
What is the best embedding model for RAG?
On Replicate, BGE-Large (nateraw/bge-large-en-v1.5) for English — BGE models lead retrieval benchmarks at this size. BGE-M3 if the corpus is multilingual.
Can I mix embedding models in one index?
No — vectors from different models live in incompatible spaces. Pick one model per index and re-embed everything if you switch.
What are sparse vs dense embeddings?
Dense vectors capture meaning in a few hundred dimensions; sparse vectors approximate keyword matching. BGE-M3 produces both, enabling hybrid search from a single model.
How do I search images with text?
Embed images and queries with CLIP (andreasjansson/clip-features) — both land in the same vector space, so nearest-neighbor search returns matching images for a text query.