Models

What is plugsky-embed-nim and when should you use it?

plugsky-embed-nim is the NIM-profile embedding model in Plugsky's 30+ model catalogue. It provides general-purpose text embeddings for RAG, semantic search and clustering with an 8K-class input limit, exposed on the OpenAI-compatible /v1/embeddings endpoint. Choose it when your pipeline is standardised around NVIDIA-style embedding models; otherwise plugsky-embed is the simpler default and the two are interchangeable at the API level.

Key facts

Model classGeneral-purpose embedding model with a NIM-style profile
Best forRAG, semantic search, clustering and recommendations on text
EndpointPOST /v1/embeddings (OpenAI-compatible)
Input limit8K-class text per request; live limits published per model
Vector dimensionPublished per model on /models — use it to size your index
Pricing tierAvailable on platform plans; see the live pricing page
CapabilitiesBatch embeddings, similarity search, clustering — no chat completions
Product statusLive

TL;DR

  • A NIM-profile embedding model for teams that name models the NVIDIA way.
  • Same OpenAI-compatible embeddings endpoint as the rest of the family.
  • Interchangeable with plugsky-embed at the API level — pick per pipeline standard.
  • Dimension lives on /models; set it once when you create the collection.
  • Peer fallback inside the embed family keeps indexing jobs alive.

How it works, step by step

  1. Read the live card at /models for dimension, input limit and status.
  2. Create your vector collection with exactly that dimension.
  3. Chunk documents to fit the 8K-class input limit, preserving section boundaries.
  4. Embed a sample corpus and test known queries to sanity-check neighbours.
  5. Benchmark retrieval quality against your current embedding model.
  6. Switch the model name in your embeddings call and re-index.
  7. Keep model and dimension recorded with the collection for future migrations.
1Read the live cardat /models fordimension, input2Create your vectorcollection withexactly that3Chunk documents tofit the 8K-classinput limit,4Embed a samplecorpus and testknown queries to5Benchmark retrievalquality againstyour current6Switch the modelname in yourembeddings call and

Try it yourself

Open the LLM cost calculator →

What plugsky-embed-nim is

plugsky-embed-nim is an embedding model for converting text into vectors used by search, RAG and clustering systems. Its distinguishing feature is its NIM-style profile: teams that already standardise on NVIDIA-style naming and tooling can adopt it without renaming models in their codebase or documentation. It serves the same OpenAI-compatible /v1/embeddings endpoint as the rest of the Plugsky catalogue.

Dimension, input limit and status are published per model at /models. Read the live value there before creating an index — the collection dimension is the one field that must match exactly.

When to choose it

Choose plugsky-embed-nim when your organisation already speaks the NIM naming convention, or when a migration plan lists NIM-style models explicitly. It is also a reasonable general-purpose choice for RAG and semantic search when you do not need multilingual behaviour.

If your corpus is English-only and naming is not a constraint, plugsky-embed is the simpler default. For mixed-language and Arabic content, use plugsky-embed-multilingual. All three live in the same family and fall back to one another automatically, but you should still pick one deliberately per collection.

Dimensions, indexing and trade-offs

Embedding models are cheap to call and expensive to migrate, so the important decisions happen at index time.

  • Fix the dimension from the live model card before you create the collection.
  • Never mix vectors from different models in one collection.
  • Chunk by structure, keep titles with passages, and store source metadata for citations.
  • Re-embed the corpus whenever you change embedding models — plan for it as a batch job.
  • Combine embeddings with keyword search for codes, names and rare tokens.

How to switch to plugsky-embed-nim

Switching is one model name: {"model": "plugsky-embed-nim", "input": ["first passage", "second passage"]} posted to https://api.plugsky.com/v1/embeddings.

Index into a fresh collection, run your labelled retrieval checks, then flip the application. Because the whole family shares one endpoint and request shape, you can A/B two embedding models against the same query set without touching application code.

Honest comparison

Dimensionplugsky-embed-nimplugsky-embedManaged vector database embeddings
Best fitNIM-standardised pipelines and general RAGDefault general-purpose embeddingsTeams that want embeddings inside their vector store
NamingNIM-style model profilePlugsky default nameVaries by vendor
EndpointOpenAI-compatible /v1/embeddingsOpenAI-compatible /v1/embeddingsVendor-specific API
DimensionRead the live value at /modelsRead the live value at /modelsVaries; may not expose a choice
MigrationChange the model name, re-indexChange the model name, re-indexOften locked to the store
FailoverAutomatic peer fallback in the embed familyAutomatic peer fallbackVendor-dependent

Frequently asked questions

Is plugsky-embed-nim free?

Embeddings are part of platform plans. The free plan includes two free chat models, plugsky-micro and plugsky-lite. See the live pricing page for plan details and trial access.

What dimension does it output?

The live output dimension is listed per model on the catalogue at /models. Use that exact value when you create the vector collection.

How is it different from plugsky-embed?

Functionally they serve the same embedding role and the same API. plugsky-embed-nim carries a NIM-style profile for teams that standardise on NVIDIA naming; plugsky-embed is the default name.

What is the input limit?

The model accepts an 8K-class text window per request. Check /models for the current live limit and size your chunks accordingly.

Can it generate text?

No. It only produces embeddings. For generation in a RAG pipeline, pair it with a chat model from the same catalogue.

Do I need to re-index when switching from another embedding model?

Yes. Embeddings from different models are not comparable, even at the same dimension. Re-embed the corpus into a new collection before switching traffic.

Does it work with any vector database?

Yes — it returns standard numeric vectors, so pgvector, Qdrant, Pinecone, Weaviate and similar stores work without special integrations.

What happens if embeddings fail mid-job?

Requests fall back to same-profile peers in the embed family, so batch indexing normally completes without intervention. Live health is on the status page.