Which Plugsky model is best for RAG?

Key facts

Embedding default	plugsky-embed for general English-dominant retrieval
Multilingual embeddings	plugsky-embed-multilingual for mixed-language corpora
Answer tiers	plugsky-lite or plugsky-micro for volume; plugsky-plus or plugsky-pro for synthesis
Endpoints	/v1/embeddings and /v1/chat/completions, OpenAI-compatible
Vector dimension	Published per model on /models — match it in your index schema
Input limit	8K-class text per embedding request; live limits are published per model
Free plan	plugsky-micro and plugsky-lite, no card required
Product status	Live

TL;DR

• RAG needs two decisions: an embedding model and an answer model.
• plugsky-embed is the default; plugsky-embed-multilingual covers mixed languages.
• Keep high-volume retrieval on the free chat tiers and escalate synthesis when needed.
• Measure retrieval recall separately from answer faithfulness.
• Re-embed the index whenever you change embedding models.

How it works, step by step

1. Pick an embedding model — plugsky-embed for one language, plugsky-embed-multilingual for mixed corpora.
2. Chunk documents on semantic boundaries and store text, metadata and chunk IDs together.
3. Build a retrieval eval set: questions with known correct source passages.
4. Score candidate answer models on grounded question answering, not generic chat.
5. Route routine questions to the free chat tiers and escalate complex synthesis.
6. Require citations in the answer format and validate them against retrieved chunks.
7. Version the index with the embedding model name so upgrades are reversible.

Try it yourself

Open the RAG model selector →

The two-model pipeline

A RAG stack has an embedding step and a generation step, and weak retrieval cannot be fixed by a better chat model. Embed with plugsky-embed for English-dominant corpora, or plugsky-embed-multilingual when documents and queries mix languages.

Evaluate each stage on its own: retrieval recall measures whether the correct passage was found at all, while answer faithfulness measures whether the model used it without inventing anything. Changing both at once makes debugging impossible.

Choosing the answer model

Generation is where tiers differ. plugsky-lite and plugsky-micro handle high-volume, short-context answers at no cost on the free plan. When answers must synthesise several retrieved passages, follow complex instructions or admit uncertainty gracefully, move to plugsky-plus or plugsky-pro. For adversarial or high-stakes domains, a top tier can act as a review pass over a cheaper draft.

• Ask for citations and a short confidence note, then validate both in code.
• Feed only the top-ranked chunks — more context is not automatically better.
• Instruct the model to answer only from context and to say when the context is insufficient.
• Log retrieved chunk IDs with every answer so failures can be replayed.

Operating the index

An embedding model is part of your schema, not a config detail. Store the vector dimension and model name with the index, and treat an embedding upgrade as a migration: build the new index alongside the old, compare retrieval quality on your eval set, then switch.

Batch embedding calls, cache vectors for unchanged documents, and re-index only what changed. The RAG API guide covers the retrieval loop end to end; pricing on self-serve plans is flat monthly with fair-use usage, so plan capacity around the live pricing page rather than per-token estimates.

Honest comparison

Choice	Option A	Option B	How to decide
Embedding model	plugsky-embed	plugsky-embed-multilingual	One language or mixed languages
Answer model	Free tiers (micro, lite)	Paid workhorses (plus, pro)	Score grounded Q&A on your eval set
Chunking	Fixed-size windows	Semantic boundaries	Measure retrieval recall
Index maintenance	Keep current model	Re-embed on model change	Version the index with the model name
Evaluation	Retrieval recall	Answer faithfulness	Track both separately

Frequently asked questions

Which Plugsky model should I use for RAG?

Use plugsky-embed for retrieval and a chat tier for synthesis. The free models, plugsky-micro and plugsky-lite, cover high-volume answers; move to plugsky-plus or plugsky-pro when answers need more reasoning or longer context.

Which embedding model should I use?

plugsky-embed is the general default. Choose plugsky-embed-multilingual when your documents and queries mix languages, and compare retrieval recall on your own corpus before switching.

What dimension are the embeddings?

The vector dimension is published per model on the live model catalogue. Store it with your index schema and re-embed everything if you change embedding models.

Can I run RAG on the free plan?

Yes — the free plan includes plugsky-micro and plugsky-lite with no card required, plus access to the embeddings endpoint. It is enough to build and evaluate a working retrieval pipeline.

How do I stop the model hallucinating?

Ground every answer in retrieved chunks, instruct the model to answer only from context, require citations, and validate those citations against the chunks you actually retrieved.

How do I evaluate the pipeline?

Separate the stages: measure retrieval recall with questions that have known source passages, then measure answer faithfulness with the chunks held fixed. Iterate on the weaker stage.

Is there a trial for the paid tiers?

Yes — the 14-day full-access trial lets you evaluate paid answer models on your own retrieval workload. See the live pricing page for current details.