Data flow 

Field buses (BACnet / Modbus / Haystack / JSON / CSV)
  → driver tree + discovery
  → Haystack model points + assignment graph
  → Arrow RecordBatches → Feather files on disk
  → DataFusion SQL rules
  → fault state, plots, PDF reports
  → React dashboard

Driver → model → FDD

  1. Drivers expose points in the driver tree (BACnet objects, Modbus registers, Haystack refs, JSON endpoints).
  2. Assignments bind driver points to Haystack semantic IDs (/api/model/assignments).
  3. Historian persists normalized samples under workspace/data/historian/.
  4. SQL rules reference semantic point names via DataFusion — not raw BACnet instance numbers.
  5. Faults surface on the dashboard and in exports/reports.

CSV path

Batch CSV loads use the headless import API (no dashboard UI):

  1. Host script or sidecar delivers CSV → POST /api/csv/import/preflight → execute
  2. Historian rows carry source_driver: csv for downstream SQL

Large CSV analysis belongs in the Pandas cookbook outside the edge.

Reports

Report builder pulls from model, historian, rules, and fault state → PDF via /api/reports/*/render/pdf.