AI Data Analysis: Query Your Database in Natural Language

Scenario overview

"How did retention change for users acquired via podcast ads last quarter?" — a great growth question that took three Slack replies, one dbt run and a Looker export to answer. NL2SQL pipelines collapse that into one prompt and one chart. The challenge is doing it without letting the model invent or leak data: schema-grounded generation + read-only credentials + result verification is what separates a demo from a production copilot.

Architecture (text version)

Code & configuration

NL2SQL with grounded schema

curl https://api.dr-ai.top/v1/chat/completions \
  -H "Authorization: Bearer sk-***" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5",
    "temperature": 0,
    "response_format": {"type": "json_object"},
    "messages": [
      {"role": "system", "content": "Translate the user question to SELECT-only SQL using the schema provided. Reply JSON: {sql, chart_type, columns}."},
      {"role": "user", "content": "Question: \"Weekly revenue by region, last 8 weeks, excluding returns\"\n\nSchema:\nCREATE TABLE orders(id, region, gross_revenue, refunded_at, created_at);"}
    ]
  }'

Sample dialogue

user > Top 5 declining SKUs by unit sales, last 30 days vs prior 30 days

assistant >
  sql
  WITH cur AS (SELECT sku, SUM(units) AS u FROM sales WHERE day >= now() - interval '30' day GROUP BY sku),
       prev AS (SELECT sku, SUM(units) AS u FROM sales WHERE day BETWEEN now() - interval '60' day AND now() - interval '30' day GROUP BY sku)
  SELECT cur.sku, cur.u AS cur_units, prev.u AS prev_units, prev.u - cur.u AS drop_units
  FROM cur JOIN prev USING(sku) ORDER BY drop_units DESC LIMIT 5;

  narrative
  SKU 78520 fell from 12,400 → 3,100 units (-75%). Confirm inventory status before reordering — this may be a stockout rather than demand drop.

Data security

• Read-only execution — the SQL runner connects with a SELECT-only role, no DDL/DML, with a 30-second hard timeout and a 50k-row cap.
• Schema, not data — only column descriptions and 5 sample rows are sent to the model. Never customer PII.
• PII redaction — outbound text is filtered through an allow-list regex; unauthorised column names (email, phone…) raise a warning before prompt is sent.
• Audit log — every (user, question, sql, row-count) tuple is logged for compliance.

ROI data

A medium-sized product org typically spends $480k/year on BI analyst headcount to answer self-serve questions. The same team, augmented with an NL2SQL copilot, supports 5× the question volume at ~$2k/month DrAI subscription cost — and analysts pivot to higher-value strategic work.

Related blog articles

Related use cases