Retrievers

The engine never touches a database. It asks your code for rows through five small interfaces (Java: async interfaces; Python: plain callables; Rust: traits, usually closures):

Interface	Signature (conceptual)	Role
EntityRetriever	(table, ids, bound) → rows	Batched point lookup: seed rows, parents
LinkRetriever	(link, parent_id, bound, limit) → rows	Children along one FK link, newest-first
CohortRetriever (optional)	(table, anchor, bound, limit) → ids	Similar entities for in-context examples
TableScanner (optional)	(table, bound) → row stream	Bulk streaming; enables FOR EACH and CSC mode
StatsProvider (optional)	—	Normalization statistics

Rows

A Row carries typed cells, an optional timestamp, and parent edges ({fk_column: parent_id}). IDs and FK values are never cells — they surface only as identity and edges.

Wiring

A RetrieverWiring binds retrievers to tables and links, with a default_links catch-all. It is validated against the schema when the engine is built, so a missing retriever fails fast, not mid-query.

Contract essentials

• Return nothing newer than the TemporalBound you are given. (The engine re-checks anyway — see Temporal correctness.)
• LinkRetriever returns children newest-first, capped at limit.
• Retrievers own their I/O: batching, caching, auth, and concurrency are yours. In Java the SPI is async (CompletionStage); Python and Rust are deliberately synchronous.