What is the Plugsky embeddings API?

Key facts

API shape	OpenAI-compatible POST /v1/embeddings
Models	plugsky-embed-v1 (1536 dimensions, ada-compatible) and plugsky-embed-large (3072 dimensions)
Batch input	Arrays of inputs accepted per request; chunk client-side for larger jobs
Use cases	Semantic search, RAG retrieval, clustering, recommendations and classification
Languages	Cross-lingual retrieval with strong support for Arabic and other languages
Vector stores	Works with pgvector, Pinecone, Qdrant or your own database
Pricing model	Embeddings are included in flat self-serve plans, with no per-vector meter
Product status	Live

TL;DR

• Drop-in OpenAI embeddings: change the base URL and keep your code.
• 1536 dimensions for general retrieval, 3072 for high-recall multilingual work.
• Embeddings are included in flat self-serve plans with no per-vector meter.
• Pair embeddings with the RAG API for end-to-end document Q&A with citations.
• Bring your own embedding model on Enterprise private endpoints.

How it works, step by step

1. Pick plugsky-embed-v1 for general retrieval or plugsky-embed-large for high recall.
2. Chunk documents with enough overlap to preserve meaning across boundaries.
3. Call /v1/embeddings with an array of inputs and store the resulting vectors.
4. Index vectors in pgvector, Pinecone, Qdrant or your own database.
5. Query with the same model used for indexing and rank by cosine similarity.
6. Feed the top chunks into a chat model for grounded answers.
7. Re-embed when you change models, because vectors from different models are not comparable.

Try it yourself

Open the embedding API tester →

Why embeddings

Embeddings turn text into a fixed-size vector where similar meaning lands at nearby points. That property powers semantic search, recommendation, clustering, RAG retrieval and anomaly detection, all without keyword matching.

Because the endpoint is OpenAI-compatible, any SDK that already speaks OpenAI embeddings works against Plugsky unchanged. You keep your pipeline and change the base URL and model name.

Plugsky's endpoint is part of a 30+ model platform, so embeddings, chat and RAG share one key, one SDK and one flat plan.

Two models, two use cases

plugsky-embed-v1 returns 1536 dimensions and is the general-purpose, OpenAI ada-compatible option for fast retrieval. plugsky-embed-large returns 3072 dimensions and targets high-recall retrieval, multilingual content and longer text.

Choose by query type, not by size alone. If recall matters more than latency and your corpus mixes languages, the larger model usually pays for itself; for high-volume English retrieval the smaller model is often enough.

Batch, storage and use cases

The endpoint accepts an array of inputs per request, so you can embed a batch of chunks in one call and write the vectors to your store. For very large jobs, chunk client-side and submit in manageable batches.

Common patterns: RAG retrieval with the RAG API, semantic search over pgvector or Pinecone, near-duplicate detection with cosine similarity, recommendation feeds, k-means topic clustering, and zero-shot classification by comparing an embedding to labelled centroids. On Enterprise, private endpoints can host your own embedding models.

Because the endpoint accepts arrays, a corpus of a few thousand chunks can be embedded in a handful of calls rather than one request per document, which keeps large ingestion jobs practical.

Honest comparison

Task	plugsky-embed-v1	plugsky-embed-large	Bring your own model
Dimensions	1536	3072	Varies
Compatibility	OpenAI ada-style	OpenAI-compatible endpoint	Varies by model
Best for	General retrieval, fast	High-recall multilingual and longer text	Specialised workloads
Vector stores	pgvector, Pinecone, Qdrant	Same	Same
Pricing	Included in flat plans	Included in flat plans	Enterprise contract
Availability	Self-serve	Self-serve	Enterprise private endpoint