How do you build translation with RAG?

Key facts

Embeddings	plugsky-embed-multilingual is live for glossary and translation-memory vectors
Generation	OpenAI-compatible chat completions with JSON mode for segment output (live)
Pipeline	Segment, embed, retrieve terms and matches, translate, verify
Models	30+ models; route by language pair and content type
Metadata	Store locale, domain, status and reviewer with every approved pair
Pricing	Flat monthly self-serve plans with no per-token charges on self-serve
Free tier	plugsky-micro and plugsky-lite on the free plan, no card required
Roadmap	The batch endpoint is coming soon; run bounded-parallel translation workers today

TL;DR

• Retrieve the glossary entry and past translations before translating each segment.
• Embed approved pairs so terminology improves as review approves them.
• Translate at segment level: cheaper to retry, easier to review.
• Keep placeholders, tags and numbers under deterministic checks.
• Store reviewer and status metadata to govern what retrieval may use.

How it works, step by step

1. Extract source content as segments with stable ids and surrounding context.
2. Build the retrieval corpus: glossary entries and approved source-target pairs with locale and domain metadata.
3. Embed the corpus with plugsky-embed-multilingual and index it for filtered search.
4. For each segment, retrieve the closest glossary terms and translation-memory matches.
5. Translate with retrieved evidence in the prompt and an instruction to preserve placeholders exactly.
6. Verify output deterministically and route flagged segments to reviewers.
7. Write approved pairs back into the corpus so the next round retrieves better matches.

Try it yourself

Open the embedding model comparison →

The retrieval corpus behind good translation

Translation quality is mostly terminology quality. Before touching the translator, build the corpus it will consult: approved glossary entries per locale and domain, plus previously approved source-target pairs from translation memory. Embed both with plugsky-embed-multilingual and store metadata — locale, domain, product, status, reviewer — so retrieval can filter to what is actually approved for this content.

The retrieval step runs per segment, not per document. A product string, a legal clause and a marketing headline need different evidence, and segment-level retrieval keeps the prompt small and relevant. Retrieve a handful of candidates, drop anything below a relevance threshold, and pass the rest with a clear instruction to use the approved term verbatim.

A segment-level pipeline with deterministic checks

Structure the pipeline so each stage is checkable:

1. Segment: split on sentence and UI-string boundaries, keeping ids stable across edits.
2. Retrieve: fetch glossary entries and translation-memory matches for the segment.
3. Translate: request JSON output per segment so parsing never depends on prose.
4. Verify: check placeholders, tags, numbers, untranslated fragments and length limits.
5. Review: send flagged segments to humans with full context.

Deterministic checks catch the failures that matter most in localization — a dropped variable or a translated product name — without spending model calls. Run them before any LLM-based quality pass so the expensive checks only see plausible candidates.

Consistency, review and measurement

Retrieval is what makes terminology consistent across a whole product over time. When reviewers approve a pair, write it back to the corpus with its metadata; the next segment that mentions the term retrieves the approved phrasing instead of reinventing it. Expire or deprecate entries rather than deleting them, so audits can reconstruct which rules were active at any point.

Measure quality per language pair and content type: terminology adherence, placeholder integrity, review edit rate and time-to-publish. Route cheap multilingual tiers for high-volume UI strings and stronger models for legal or brand-sensitive copy. The free plan with plugsky-micro and plugsky-lite is enough to build the pipeline, and the 14-day full-access trial lets you benchmark stronger models per pair before committing.

Honest comparison

Concern	RAG translation on Plugsky	Prompt-only translation	Rules-based localization
Terminology	Retrieved from approved glossary	Recall-dependent	Explicit term rules
Reuse	Translation memory via embeddings	Manual copy-paste	Segment matching
Segment control	Per-segment output with ids	Whole-document pass	Segment engine
Adaptation	New approved pairs improve results	Prompt edits	Rule maintenance
Coverage	30+ models for any language pair	One model per prompt	Engine-dependent

Frequently asked questions

What does RAG add to translation?

It supplies the terminology and prior approved translations for each segment, so the model translates with your vocabulary instead of inventing one. Consistency improves as the retrieval corpus grows.

Which embedding model should I use?

plugsky-embed-multilingual is built for mixed-language corpora, which is exactly what glossaries and parallel translation pairs are. Use the same model for the corpus and the queries.

How do I keep placeholders safe?

Request structured output per segment and run deterministic checks for variables, tags and numbers. Any mismatch goes to review, not to production.

Can I reuse previous translations?

Yes. Store approved pairs in translation memory, retrieve the closest matches with embeddings, and instruct the model to prefer them when they fit the context.

How do I handle brand names?

Mark them as do-not-translate in the glossary, retrieve them for every segment where they appear, and add a verification check that fails if the name changed.

Does this work for right-to-left languages?

Yes, the pipeline is language-agnostic: retrieval, translation and checks work the same way. Evaluate quality per pair, because model strength varies by language direction.

Is large-scale batch translation available?

The batch endpoint is coming soon. Today, process segments in bounded-parallel workers with retries and a queue, and keep review throughput as the real constraint.

Can I start for free?

Yes. plugsky-micro and plugsky-lite are on the free plan with no card, and a 14-day full-access trial is available for stronger multilingual models.