Skip to content
MountSQLiMountSQLI

AI Overview

@mountsqli/ai is AI native by design: generated queries go through the same compiler validator as hand-written ones. There is no separate, unsafe SQL generator.

Capabilities

Section titled “Capabilities”
Function What it does
nlToSql natural language → QueryPlan (not raw SQL)
explainPlan human-readable explanation of a plan
optimizePlan suggests optimizations for a plan
reviewPlans security/perf findings across plans

The ModelProvider seam

Section titled “The ModelProvider seam”

The only external dependency is ModelProvider — an interface you implement (OpenAI, Anthropic, a local model, or a test fake). The package never calls a vendor SDK directly.

import { Ai } from "@mountsqli/ai";


const ai = new Ai({ provider: myProvider });
const result = await ai.nlToSql("users over 18", [users.def]);
// result.plan is a QueryPlan — compiled & type-checked like any other

Why plans, not strings?

Section titled “Why plans, not strings?”

Because the model emits a QueryPlan, the result is:

  • Type-checked against your schema.
  • Injection-safe — values are parameters, never concatenated.
  • Explainable — explainPlan can describe it.

Best practices

Section titled “Best practices”
  • Validate the returned plan (compile it) before running, like any query.
  • Swap ModelProvider per environment (fake in tests).
  • Use reviewPlans in CI to catch risky generated queries.

Common mistakes

Section titled “Common mistakes”
  • Treating model output as trusted SQL — still compile the plan.
  • Hard-coding a vendor SDK in app code (use ModelProvider).
Section titled “Related”