Does Plugsky support RAG?

Key facts

RAG status	Live — embeddings plus chat completions
Embeddings	/v1/embeddings with the plugsky-embed family
Multilingual	plugsky-embed-multilingual for mixed-language retrieval
Generation	Any chat model; cheap aliases for drafting
Vector store	Bring your own — no managed vector database today
Reranking	Not a documented live endpoint
Batch ingestion	Coming soon
Free tier	2 free AI models (plugsky-micro, plugsky-lite), no card

TL;DR

• RAG is live: embeddings for retrieval, chat for grounded generation.
• One OpenAI-compatible API serves both halves of the pipeline.
• Your vector store, chunking and access control stay yours.
• Reranking and batch ingestion are not in the live stack yet — plan accordingly.
• Evaluate retrieval separately from generation to find real failures.

How it works, step by step

1. Chunk documents with structure in mind and attach stable ids plus provenance metadata.
2. Embed chunks with plugsky-embed (or plugsky-embed-multilingual) and store vectors in your database.
3. Embed the query, retrieve top-k candidates, and apply your own metadata filters.
4. Compose an answer with mandatory citations to the retrieved passages.
5. Evaluate retrieval recall and answer faithfulness separately on a labelled set.
6. Monitor index freshness and re-embed when documents change.

Try it yourself

Open the RAG sandbox →

What Plugsky provides for RAG

Plugsky covers the model layer of a retrieval pipeline:

• Indexing and retrieval: POST /v1/embeddings turns chunks and queries into vectors, with plugsky-embed-multilingual available when documents mix languages such as Arabic and English.
• Generation: any chat model can compose the final answer, and cheap aliases handle sub-query rewriting and drafting while a stronger model handles synthesis.
• Grounding controls: function calling lets retrieval become a tool the model calls, which suits agentic patterns and multi-hop questions.

Everything is OpenAI-compatible, so libraries such as LangChain or LlamaIndex work with a base URL change — see the integration guides in the docs.

What stays your responsibility

A RAG system is more than two API calls, and the parts Plugsky does not own decide your answer quality:

• Chunking: sizes, overlap and structure preservation.
• Vector store: hosting, indexing strategy, metadata filters and backups.
• Authorisation: ensuring a retrieval result is content the requesting user may see.
• Freshness: re-indexing changed or deleted documents.
• Evaluation: labelled questions for recall and faithfulness.

Reranking is not a documented live endpoint and batch ingestion is coming soon, so build those stages yourself if you need them now.

What we do and what we do not do

What we do: provide embeddings and chat models behind one API, support multilingual retrieval, and keep the interface compatible with the RAG frameworks teams already use. What we do not do: host your vector database, guarantee retrieval quality, or take over chunking and permissions — those are application design choices with big quality consequences. If you want a managed end-to-end knowledge product, that is a different class of tool. For a composable pipeline you control, the building blocks are live today.

Honest comparison

Layer	Plugsky	Managed RAG platform	Building in-house only
Embeddings	Live API	Included	You host a model
Chat generation	30+ models	Included	You host a model
Vector store	Bring your own	Managed	You build
Reranking	Not in live stack	Often included	You integrate
Access control	Your application	Platform features	You build
Operational effort	Moderate — you own the pipeline	Low	High

Frequently asked questions

Do I need a separate embedding provider?

No. Plugsky serves embeddings through /v1/embeddings, so retrieval and generation can use the same API key and base URL.

Which embedding model should I start with?

plugsky-embed for single-language content and plugsky-embed-multilingual when documents mix languages such as Arabic and English. Test retrieval precision on your own corpus.

Does Plugsky store my vectors?

No. You host the vector store; Plugsky returns vectors and does not manage your index.

Can I use LangChain or LlamaIndex?

Yes. The API is OpenAI-compatible, so these frameworks work by changing the base URL; integration guides are in the docs.

Is reranking available?

Not as a documented live endpoint today. Add a reranking stage yourself or design retrieval to compensate with better chunking and metadata filters.

How do I evaluate RAG quality?

Separate the two stages: measure retrieval recall on labelled questions, then measure answer faithfulness against the passages that were actually retrieved.

Can I prototype on the free plan?

You can prototype chat-driven retrieval with the two free models, and the 14-day full-access trial lets you compare stronger generation models. Embedding availability is documented on the pricing page.