Updated
AI models and media

Embeddings

Turn text (and sometimes images) into vectors so you can search, cluster and retrieve by meaning.

Ask your AI about this, with this page as the source:ChatGPT ↗Claude ↗Perplexity ↗

Trends

Gemini API gained the most ground in package downloads: 4.9× against the median npm package over a year.

Public counts each month: pull requests on GitHub, package downloads, stars, Product Hunt followers. They show direction, not market share.

ToolDownloadsStarsProduct Hunt
OpenAI API37.4M/wk+60% vs npm—34.5k followers
Gemini API22.2M/wk4.9× vs npm—7.7k followers
Sentence Transformers5.3M/wk★ 19.2k—
Cohere465k/wk−57% vs npm——
Voyage AI244.5k/wk+111% vs npm——

The real choice

Stay with the vendor you already use, pick a retrieval specialist, or run an open model yourself. Using your main LLM provider's embeddings means one key and one bill. Specialists offer models tuned for retrieval, many languages or specific domains, plus rerankers, at the cost of one more vendor. Open embedding models run on your own hardware for free and keep text private, but you host and scale them.

What makers chose14 live products, each with a source. Not market share.

Pick by situation

Tap the ones that are you — the tools that fit light up below.
  • You already use OpenAI or Gemini for text and want no new vendorOpenAI APIGemini API
  • Your users search in many languages, or you want embeddings and a reranker from one APICohereJina AI
  • Generic embeddings miss the terms in your code, legal or finance documentsVoyage AI
  • Text must not leave your servers, or embedding volume makes per-token pricing add upSentence Transformers

Compare the tools

Grouped by the side of the choice they answer, not ranked. Open a row for when to use it, the trade-off and what makers say.

ToolBest for
General LLM provider2
Gemini APIapi · free tierApps on the Gemini API that want multilingual text embeddings on the same key.No evidence yet22.2M/wk4.9× vs npm

Use it whenGemini is already your main model provider.

Trade-offGeneral-purpose models, and like any hosted embedding, switching later means re-embedding everything.

In Lovable, Replit

No maker's product we track shows it for this yet.

OpenAI APIapiApps already on the OpenAI API that want solid embeddings with no new vendor.No evidence yet37.4M/wk+60% vs npm

Use it whenYou already use OpenAI for text and your content is general-purpose English or common languages.

Trade-offGeneral-purpose models, with no reranking endpoint to pair them with.

llms.txtIn Lovable, Replit, Rork

No maker's product we track shows it for this yet.

Retrieval specialist3
CohereapiMultilingual search with embeddings and a reranker from the same vendor.19+58 open source465k/wk−57% vs npm

Use it whenYour users search in many languages or you want reranking in one API.

Trade-offFree trial keys are rate-limited and not for commercial use.

Used by Chat Data, CustomGPT, Internet.io, Mnemosphere AI, Officely AI and 14 more · in 58 open-source projects.

We use Cohere as an endpoint in our team builder flow; you can combine this model with all available models.
Officely AI, the makerSep 2026 ↗
Voyage AIapiRetrieval quality on specialized text, with models tuned for code, legal and finance documents.22 open source244.5k/wk+111% vs npm

Use it whenGeneric embeddings miss domain terms in your data.

Trade-offHosted API apart from one small open-weight model; now run by MongoDB, so its roadmap follows MongoDB's.

No maker's product we track shows it for this yet · in 22 open-source projects.

Jina AIapiMultilingual and long-document embeddings you can try without an API key.No evidence yet

Use it whenYou also need its Reader or reranker and want one vendor for the retrieval pipeline.

Trade-offKeyless use is rate-limited; now part of Elastic, so the product direction may shift.

No maker's product we track shows it for this yet.

Open model you run1
Sentence Transformerslibrary · open sourceRunning open embedding and reranker models on your own CPU or GPU, with thousands of models on the Hugging Face Hub to choose from.125 open source5.3M/wk

Use it whenDocuments are private, volume is high, or you want to fine-tune an embedding model on your own data.

Trade-offPython only; you host, batch and scale the model yourself, and CPU-only servers embed slowly.

No maker's product we track shows it for this yet · in 125 open-source projects.

Who switches to what

Public pull requests on GitHub since Oct 2024 whose title says "X to Y" — real code changes, by developers in general rather than makers only. Pick a flow to see its pull requests.

OpenAI API → Gemini API178+ PRs

Before you choose

How to approach it

Pick a model by testing it on 50–100 real queries from your own data, not by public leaderboards. Record the model name and dimensions with every stored vector — switching models later means re-embedding everything. Adding a reranker on top of the first search often improves results more than changing the embedding model.

Common mistakes
  • Mixing vectors from two embedding models, or two dimensions of one model, in the same index; similarity scores between them are meaningless.
  • Embedding whole documents as one vector, so a search can only ever return "the whole PDF" instead of the passage that answers the question.
  • Choosing a model from a public leaderboard without testing it on your own queries and documents.

Decided alongside

What the 14 makers' products here chose for their other decisions.