// Module 01

Conversational Data Copilot

Ask your MES in plain language. Get correct SQL, every time.

An agentic Text-to-SQL engine grounded in a per-site domain dictionary. Operators speak Swedish, English or Chinese; the appliance returns scoped, JOIN-correct T-SQL and a human answer — without ever touching the live SQL Server.

Inputs

Natural language: Free-form question (SV / EN / ZH)
Schema dictionary: Per-site RAG JSON: tables, columns, synonyms, business terms
Routing rules: Which question types map to which table families (MES / VDS / mistake-proof)
Session context: Last 5 turns + selected line / station filter

Outputs

Generated T-SQL: Constrained, syntactically valid, scoped to allowed schemas
Result table: Read against the air-gapped sandbox, never the live DB
Plain-language answer: Same language as the question, with units and timestamps
Why this query: One-line trace: dictionary entries + routing rule that fired

Example prompts

Operator

What was OEE on the production line yesterday, and what caused most downtime?

Generated SQL

SELECT TOP 1 cause, SUM(duration_s) AS total
FROM alarm_log
WHERE line = 'LINE01'
  AND start_ts >= DATEADD(day, -1, CAST(GETDATE() AS date))
  AND start_ts <  CAST(GETDATE() AS date)
GROUP BY cause
ORDER BY total DESC;

Appliance

OEE on the line yesterday was 78.4%. The largest downtime bucket was material starvation at station Station 03 (42 min across 3 events).

Operator

Vilket arbetsorder hade långsammast takt idag?

Generated SQL

SELECT TOP 5 work_order, AVG(cycle_time_s) AS avg_cycle
FROM cycle_log
WHERE shift_date = CAST(GETDATE() AS date)
GROUP BY work_order
ORDER BY avg_cycle DESC;

Appliance

En arbetsorder hade dagens långsammaste takt: medel 62 s mot målet 48 s. Avvikelsen kommer från Station 02.

Operator

Show VDS hardware diagnostics for the tightening tool in the last two hours.

Generated SQL

SELECT ts, event_code, severity, message
FROM device_diag_log
WHERE device_id = 'TOOL-01'
  AND ts >= DATEADD(hour, -2, GETDATE())
ORDER BY ts DESC;

Appliance

47 events in the last 2 h on the tightening tool. 11 are network heartbeat drops clustered between 13:42–13:51; the remaining 36 are routine torque confirmations.

Operating notes

[01]All execution happens against the sandbox copy of SQL Server. The live MES database is never reached, never written to.
[02]Generated SQL is constrained to schemas listed in the domain dictionary — out-of-scope tables cannot be queried.
[03]If the dictionary is missing a term, the appliance asks a single clarifying question rather than hallucinating a JOIN.

Other modules

Auto-Reporting Engine

Shift reports and root-cause briefs, written by the appliance.

Predictive Sentinel

Warnings minutes-to-hours before the line actually stops.