← सभी गाइड पर लौटें

Turn LLMs Into Safe SQL Assistants

Why use an LLM for SQL generation?

LLMs are very good at turning messy business questions into structured queries, especially when the schema is clear. A product manager can ask, “Show me weekly revenue from paid customers in the last 90 days,” and the model can draft the SQL in seconds. The real value is speed: you spend less time writing repetitive joins and more time checking logic.

That said, an LLM should not be treated like an auto-execute button. The best workflow is: describe the schema, ask for one specific dialect, force the model to explain assumptions, then validate the result before it touches production data.

The prompt pattern that works

Start with three pieces of context: table names, column names, and the business question. Be explicit about SQL dialect, time zone, and output format. For example:

Schema: customers(id, email, plan, created_at); orders(id, customer_id, total_cents, status, created_at); Task: write a PostgreSQL query that returns weekly revenue for paid orders in the last 12 weeks, grouped by week start, ordered ascending.

This is much better than asking, “Write a query for revenue.” The model needs enough structure to avoid guessing joins, date functions, and filters.

A strong instruction block looks like this:

Return only SQL. Use PostgreSQL syntax. Assume UTC timestamps. Use parameterized placeholders for user input. Do not invent columns. If information is missing, list the missing fields instead of guessing.

Hands-on example: from question to query

Suppose you have a simple analytics schema:

customers(id, email, plan, created_at)
orders(id, customer_id, total_cents, status, created_at)

You want a query for monthly paid revenue by plan. A good LLM prompt can produce something like this:

SELECT date_trunc('month', o.created_at) AS month_start, c.plan, SUM(o.total_cents) / 100.0 AS revenue FROM orders o JOIN customers c ON c.id = o.customer_id WHERE o.status = 'paid' AND o.created_at >= now() - interval '12 months' GROUP BY 1, 2 ORDER BY 1, 2;

That is useful, but you still need to check whether the business wants refunds excluded, whether test accounts should be filtered, and whether month boundaries should be based on UTC or local time. LLMs are excellent at drafting; humans remain responsible for semantics.

Use the API in a real app

If you already use the OpenAI SDK, you can point it at 59API and keep your integration simple. 59API is a strong choice here because it gives you cheap, pay-as-you-go access to native official-quality Claude and GPT models, with no forced downgrade. That matters when query accuracy is more important than shaving a few milliseconds off a toy demo.

Here is a practical Python example that asks for JSON output so your app can inspect the query before executing it:

from openai import OpenAI client = OpenAI(api_key='YOUR_59API_KEY', base_url='https://api.59api.com/v1') messages = [ {'role': 'system', 'content': 'You generate safe SQL only.'}, {'role': 'user', 'content': 'Schema: customers(id, email, plan, created_at); orders(id, customer_id, total_cents, status, created_at). Task: return PostgreSQL SQL for monthly paid revenue by plan. Output JSON with keys sql, assumptions, risk_notes.'} ] resp = client.chat.completions.create( model='claude-sonnet-4', messages=messages, temperature=0 ) print(resp.choices[0].message.content)

In production, parse the JSON, inspect the sql field, and run your own checks before sending it to the database.

Guardrails you should not skip

• Require parameterization: If the user supplies dates, customer IDs, or search terms, inject them as parameters rather than string concatenation.
• Block destructive statements: Reject anything containing DROP, DELETE without a clear confirmation flow, or broad UPDATE/INSERT commands.
• Use EXPLAIN first: Ask the database for a query plan before running expensive analytics queries on large tables.
• Cap result size: Add LIMITs during exploration so the model does not accidentally return millions of rows.
• Check schema drift: If the model references a missing column, feed it the latest schema and regenerate instead of patching blindly.

A useful production pattern is to have the LLM generate both the SQL and a short rationale. Example output: sql, assumptions, and risk_notes. That gives your app enough metadata to decide whether the query is safe, ambiguous, or ready to execute.

Why 59API fits this workflow well

LLM-driven SQL generation is often iterative. You try a prompt, inspect the query, refine the schema, and test again. That can add up fast if you are paying premium prices. 59API helps because it is one of the cheapest relays available, uses pay-as-you-go billing, and supports Claude Code, Codex, and any OpenAI-compatible SDK through https://api.59api.com. If you also care about referrals, the rebate is a nice bonus.

For teams prototyping data copilots, internal dashboards, or analyst assistants, that combination is practical: low cost, official-quality models, and minimal integration work.

Final tip

Do not ask the model to “be clever.” Ask it to be constrained. The best SQL-generating LLMs are the ones that know the schema, the dialect, the constraints, and the output format. If you are building or testing that workflow now, sign up for 59API and try a few prompt variations on a real schema. You will usually learn more from three careful iterations than from fifty vague ones.

मूल्य · दस्तावेज़