Skip to main content

Architecture

Every query runs through the same four stages, in every language:

 RelQL string
    │  parse       → typed AST (syntax errors here)
    ▼
 ParsedQuery
    │  validate    → bind names/types/windows against the Schema; infer TaskType
    ▼
 ValidatedQuery
    │  assemble    → hop loop through YOUR retrievers, bounded by anchor time
    ▼
 per-entity contexts
    │  score       → ModelBackend routed by TaskType
    ▼
 PredictionResult  → one prediction per entity (value and/or probability)

Inputs​

  • Query — a RelQL string. See the RelQL docs.
  • Anchor time — the "as of" instant t₀. Context may only contain data at or before it; the prediction concerns the window after it.
  • Entities — an explicit ID list, or FOR EACH over the whole table (enumerated via a TableScanner).

The schema carries shape only​

Tables, typed columns (NUMBER | TEXT | DATETIME | BOOLEAN), primary keys, per-table time columns, and FK links. No URLs, no credentials.

Validation enforces that links resolve, link targets have primary keys, and that PK/FK columns are never feature columns — IDs are graph edges, not values. There is no way to hand the model an identifier as a feature.

Execution is GraphQL-style​

The engine owns the language, planning, context assembly, and model routing. It never connects to a database: all data access goes through retrievers you implement. The same query runs against JDBC, a REST service, a feature store, or an in-memory test double — only the wiring changes.