Application data
One record, read whole: its rows, its documents, and the history behind them.
One request reads rows, keys and joins alongside nested documents and the change history behind them, instead of loading one side into the other.
- Rows for core records
- Documents for changing shape
- History read beside both
Data Infrastructure → Real-time data
What is happening now, to which record, and what led here.
Telemetry and events are stored with the records and documents they describe, so a window, the entity behind it and its context are read from one layer.
- Events and measurements
- Windowed reads with the record attached
- No export between systems
Real-time Analytics → Operational reporting
Today's number, asked of live operations — filtered by the fields that don't fit columns.
Aggregates and windows run over the live tables with document fields filterable beside them, so the number you report and the number you serve come from one place.
- Aggregates in place
- Document fields filterable
- One copy of the truth
Operational Analytics → Search and retrieval
Find what is similar — ranked, filtered by structure, and current as of the last write.
Embeddings are stored beside the records and documents they describe and ranked with distance ordering, filtered by structure, so similarity is an access path over the data that already serves the application.
- Dense vector columns
- Exact top-k, HNSW at scale
- Metadata filtering beside ranking
AI Retrieval → Connected data
What is reachable, and what depends on what.
Questions about relationships rather than records: expressing edges directly removes the reconstruction at query time.
- Nodes and edges as structures
- Traversal without rebuild
- Over data already in the layer
Knowledge Systems → Large objects
Files, media and artifacts that need to be found by query.
Objects are usually stored somewhere and described somewhere else. Keeping both in the layer means an asset is found by a query, not a convention.
- Big immutable objects
- Queryable metadata
- No separate catalogue
Document Intelligence →