Translation graph

Content-addressable, by design.

Same content, same fingerprint, translated once. A graph that's invariant to where your content lives.

What content-addressable means

The fingerprint is the key.

Every translatable unit — a paragraph, a button label, a JSON-LD property — gets a 12-character fingerprint derived from blake3(unit_type || content). The fingerprint is the address. Look it up; get the translations.

Same content always produces the same fingerprint. So if "Add to cart" appears on 500 products, it has one fingerprint — and one translation per locale, shared across all 500.

Source unit · fingerprint A3F5C9E27K2P

<button data-pg="A3F5C9E27K2P">Add to cart</button>

→ rendered as Añadir al carrito at /es/

Three consequences

What you get from this design.

Trivial dedupe

Repeated content is translated once. No background dedupe job to maintain; it's a property of the data model.

Trivial invalidation

Edit a paragraph, its content hash changes, its fingerprint changes, only that unit re-translates. Cache is content-derived.

Trivial CMS migration

The graph isn't tied to a CMS schema. Move WordPress → Sanity → Payload; the same fingerprints match the same content.

What's in the graph

The shape, in SQL.

01-data-model.md (excerpt) sql

-- One row per unique source unit per project
CREATE TABLE source_units (
  id              uuid PRIMARY KEY DEFAULT gen_random_uuid(),
  project_id      uuid NOT NULL REFERENCES projects(id) ON DELETE CASCADE,
  unit_type       text NOT NULL,           -- 'paragraph' | 'button' | 'product_name' | ...
  fingerprint     text NOT NULL,           -- blake3(unit_type || content), 12-char prefix
  content         text NOT NULL,
  embedding       vector(3072),            -- OpenAI text-embedding-3-large
  word_count      int  NOT NULL,
  created_at      timestamptz NOT NULL DEFAULT now(),
  UNIQUE (project_id, fingerprint)
);

-- Per source_unit × locale × branch
CREATE TABLE translations (
  id              uuid PRIMARY KEY DEFAULT gen_random_uuid(),
  source_unit_id  uuid NOT NULL REFERENCES source_units(id) ON DELETE CASCADE,
  locale          text NOT NULL,
  branch_id       uuid REFERENCES branches(id),
  content         text NOT NULL,
  model           text NOT NULL,           -- 'deepl-pro' | 'claude-sonnet' | ...
  quality_score   numeric(3,2),            -- 0.00–1.00 from Haiku judge
  status          text NOT NULL DEFAULT 'pending',
                                            -- 'pending' | 'approved' | 'rejected' | 'tm_match'
  approved_by     uuid REFERENCES users(id),
  approved_at     timestamptz,
  created_at      timestamptz NOT NULL DEFAULT now(),
  UNIQUE (source_unit_id, locale, branch_id)
);

One real example

A button on a product page.

1. A theme template renders <button>Add to cart</button>.
2. On save, the WordPress plugin computes fingerprint = blake3("button" || "Add to cart") = A3F5C9E27K2P. Inserts (idempotently) into source_units.
3. The translation engine looks up TM. "Add to cart" exists. translations get inserted with status='tm_match' and cost=0.
4. The theme template renders with a data-pg attribute. The Worker substitutes at the edge per locale.

Want the full data model?

Open spec — published with the marketing site, not behind a sales call.

Read the data model Talk to sales