tech-studies

Database Queries

2 min read

Database Queries

A tool that lets an agent read (and sometimes write) structured data by issuing queries — typically text-to-SQL — against a real database.

Why it matters

Most business truth lives in relational tables, not documents. An agent that can query a DB answers “what was Q3 revenue by region?” with exact numbers instead of a guess. This is the backbone of analytics copilots and operational agents. Unlike rag-and-vector-databases (fuzzy, semantic), SQL is precise and aggregatable — the right tool when the answer is a computed fact, not a relevant passage. See databases.

How it works

The model is given the schema (tables, columns, types, a few example rows), generates SQL, the tool executes it, and rows return as the observation. The schema is the single most important input — without it the model invents column names.

  • Schema injection — put CREATE TABLE statements or a compact column list in the prompt. For big schemas, retrieve only the relevant tables first.
  • Read-only by default — connect as a role with SELECT-only grants. Let the database, not a regex, enforce safety.
  • Guardrails — auto-append LIMIT, set a statement timeout, block DROP/DELETE/UPDATE unless explicitly enabled with confirmation (human-in-the-loop-evaluation).
  • Parameterize — never string-concat user values into SQL; the model writing SQL is fine, but injected parameters must be bound. See databases.
  • Self-correction — return DB errors verbatim so the model fixes a typo’d column or bad JOIN and retries.
RiskMitigation
Destructive writeread-only DB role
Full-table scanforced LIMIT + timeout
Wrong column/JOINschema in prompt + error feedback
Huge result setcap rows, summarize

Example

schema: orders(id, region, amount, created_at)
user: "total revenue per region in Q3 2025"
model → SELECT region, SUM(amount) AS rev
        FROM orders
        WHERE created_at >= '2025-07-01' AND created_at < '2025-10-01'
        GROUP BY region ORDER BY rev DESC;
rows  → [("EMEA", 412000), ("AMER", 388500), ...]

Exact aggregates a vector store could never produce.

Pitfalls

  • Hallucinated columns when the schema is missing or stale — the query 500s or, worse, silently returns wrong data.
  • Returning thousands of rows floods context-windows; aggregate or paginate, and prefer COUNT/SUM over dumping rows.
  • Semantic-not-syntactic errors — valid SQL with the wrong JOIN gives a confidently wrong number; validate against known totals.
  • Trusting “read-only” prompts — only a restricted DB grant truly prevents writes; never rely on instructions alone.

See also

Graph View

Backlinks