Skip to content
Pollenix Docs

DB workbench

Workbench

The DB Engine’s chat surface. Lives at /db.

Asking a question

  1. Pick a connection from the rail (only granted ones appear).
  2. Type a question in plain English: “average days-in-A/R by payer in Q1 2026”.
  3. Pollen8:
    • Embeds your question.
    • Retrieves the top-K most-relevant tables from the connection’s introspected schema.
    • Asks the planner LLM to write SQL that answers the question using only those tables.
    • Validates the SQL through sqlglot — SELECT-only, no dangerous statements, row-cap injected, scope predicates AND’d in.
    • Runs EXPLAIN preflight to estimate cost; blocks queries that would scan billions of rows.
    • Executes via the driver. Statement timeout 30s.
  4. Surface shows: the SQL, the result table, and a chart spec inferred from the answer shape.

Three result views

  • Table — the rows as returned.
  • Chart — auto-picked from the data shape (line if there’s a time axis, bar if there’s a category column + a metric, donut for simple breakdowns, scatter for two metrics).
  • SQL — the generated SQL. Editable — tweak and re-run.

Edit the SQL

Click into the SQL panel → the generated SQL becomes editable. Hit run → goes through the same validator + EXPLAIN preflight as the planner-generated path. You can’t sneak past the validator by hand-editing.

History

Every workbench query persists to a per-user history drawer. Click any prior question to re-open it with the SQL + chart spec.

Send to BI

Click Send to BI on any result to materialize it as a widget on a BI dashboard. The widget carries the connection id + the SQL + the chart spec; refreshing the widget re-runs the SQL.

Common patterns

  • Time-series: “Daily new patients last 90 days” → line chart.
  • Breakdown: “Encounters by provider this quarter” → bar.
  • Cohort: “Patients with A1c > 9 in last visit” → table you can send to BI as a watchlist.
  • Ratio: “First-pass rate by payer” → bar + KPI tile if you also Send to BI.

When NL→SQL fails

The planner returns “I couldn’t write SQL for that” when:

  • The question is too vague (“how are we doing?”).
  • The relevant tables aren’t introspected — re-introspect the schema.
  • The question is structurally outside DB scope (cross-connection JOINs, write operations).

Re-phrase or switch the connection.