Key facts
| Default model | plugsky-embed for general-purpose retrieval and clustering |
| NIM profile | plugsky-embed-nim for NIM-standardised pipelines |
| Multilingual | plugsky-embed-multilingual for mixed-language corpora |
| Endpoint | POST /v1/embeddings (OpenAI-compatible) |
| Vector dimension | Published per model on /models — set it in your index schema |
| Input limit | 8K-class text per request; live limits are published per model |
| Pricing tier | Embeddings are available on platform plans — see the live pricing page |
| Product status | Live |
TL;DR
- Three embedding models, one OpenAI-compatible endpoint.
- plugsky-embed is the default for general retrieval.
- plugsky-embed-multilingual is the pick for mixed-language corpora.
- plugsky-embed-nim suits pipelines that name models the NVIDIA way.
- Declare the vector dimension in your index schema before going live.
How it works, step by step
- List the languages and domains your corpus actually contains.
- Build a retrieval eval set: queries with known relevant passages.
- Score plugsky-embed and plugsky-embed-multilingual on that set.
- Include plugsky-embed-nim if your platform standardises on NIM-style naming.
- Record the winning model's vector dimension and create the index schema.
- Batch and cache embeddings, and re-embed the corpus whenever you change models.
Try it yourself
Open the embedding model comparison →
The three embedding models
plugsky-embed is the general-purpose default for semantic search, RAG and clustering. plugsky-embed-nim targets teams that standardise on NIM-profile model naming, and plugsky-embed-multilingual is built for cross-language retrieval where documents and queries may not share a language.
All three accept text through the same /v1/embeddings shape, so switching is a model-name change plus a re-embed of the index — not an integration project.
How to choose
Language mix is the first filter. An English-dominant corpus can use the default; a mixed Arabic and English corpus should be tested against the multilingual model, with the default kept as a baseline. If your organisation names models after NIM upstreams, plugsky-embed-nim keeps that convention without changing the API.
- Score on your own queries — embedding quality is domain-specific.
- Compare top-k recall, not average similarity scores.
- Test the text you really embed: short queries, long passages, tables and code.
- Re-run the comparison when the model card changes.
Record the winner and the date next to the index schema; the next engineer should not have to rediscover why that model was chosen.
Index operations that save time later
Embedding models are part of your schema. Store the model name and vector dimension with the index, and treat an upgrade as a migration: build the new index alongside the old one, compare retrieval quality, then switch. Keep the old index until you are confident enough to delete it.
Batch embedding requests where possible, cache vectors for unchanged documents, and re-index only what changed. Because pricing is flat monthly with fair-use usage on self-serve plans, capacity questions are about plan limits rather than token arithmetic — see the live pricing page for current plans.
Honest comparison
| Dimension | plugsky-embed | plugsky-embed-nim | plugsky-embed-multilingual |
|---|---|---|---|
| Default use | General retrieval, RAG, clustering | NIM-standardised pipelines | Mixed-language corpora |
| API endpoint | Same /v1/embeddings | Same /v1/embeddings | Same /v1/embeddings |
| Vector dimension | See the live model card | See the live model card | See the live model card |
| Best fit | English-dominant data | Teams using NIM naming | Arabic plus English content |
| Migration | Model-name change plus re-embed | Model-name change plus re-embed | Model-name change plus re-embed |
Frequently asked questions
Which Plugsky embedding model should I use?
Use plugsky-embed as the general default, plugsky-embed-multilingual when your corpus mixes languages, and plugsky-embed-nim when your platform standardises on NIM-style naming. Compare retrieval recall on your own queries.
Do all three share the same endpoint?
Yes — all three are served through the OpenAI-compatible /v1/embeddings endpoint, so switching models is a model-name change for existing embedding code.
What vector dimension do they return?
The dimension is published per model on the live catalogue. Check it before creating your index schema, because changing embedding models usually means re-embedding the corpus.
Can I mix embedding models in one index?
No — vectors from different models are not comparable. Use one embedding model per index, or maintain separate indexes and merge results at the application layer.
How do I test an embedding model?
Build a set of queries with known relevant passages and measure top-k recall for each candidate. Include realistic text: short questions, long passages, tables and code snippets.
Are embeddings available on the free plan?
The free plan includes plugsky-micro and plugsky-lite with no card required, and embeddings are available on platform plans. See the live pricing page for current plan details.
How is pricing structured?
Self-serve plans are flat monthly with fair-use usage rather than per-token billing. Check the live pricing page for current plans and limits.