SQL: the data skill
every AI engineer needs.

Why Bother

Where SQL shows up in AI work

The Mental Model

A query is a question about tables

A database is a set of tables (like spreadsheets): rows are records, columns are fields. All examples below use two tables you might have in a real AI project: evals(id, model, question, score, latency_ms, created_at) and models(name, provider, cost_per_1k).

The Crash Course

Eight topics = 90% of daily SQL

-- The 10 worst-scoring answers from claude models this week
SELECT question, score, latency_ms
FROM evals
WHERE model LIKE 'claude%'
  AND created_at >= '2026-07-28'
ORDER BY score ASC
LIMIT 10;

-- Average score and volume per model, best first — models with 50+ runs only
SELECT model,
       COUNT(*)            AS runs,
       ROUND(AVG(score), 2) AS avg_score
FROM evals
GROUP BY model
HAVING COUNT(*) >= 50
ORDER BY avg_score DESC;

-- Add provider and cost info to every eval row
SELECT e.model, m.provider, e.score, m.cost_per_1k
FROM evals e
JOIN models m ON e.model = m.name;

-- LEFT JOIN: keep ALL evals, even if the model isn't in the models table
SELECT e.model, m.provider
FROM evals e
LEFT JOIN models m ON e.model = m.name
WHERE m.name IS NULL;   -- finds evals with no matching model — a classic data check

-- Which models beat the overall average?
WITH per_model AS (
  SELECT model, AVG(score) AS avg_score
  FROM evals
  GROUP BY model
),
overall AS (
  SELECT AVG(score) AS avg_all FROM evals
)
SELECT p.model, p.avg_score
FROM per_model p, overall o
WHERE p.avg_score > o.avg_all;

-- Best answer per model (rank within each model's rows)
SELECT model, question, score,
       ROW_NUMBER() OVER (PARTITION BY model ORDER BY score DESC) AS rank
FROM evals;
-- wrap it in a CTE and keep WHERE rank = 1 → "top answer per model"

-- Did latency improve vs the previous run?
SELECT created_at, latency_ms,
       latency_ms - LAG(latency_ms) OVER (ORDER BY created_at) AS delta
FROM evals WHERE model = 'claude-sonnet-5';

CREATE TABLE evals (
  id         SERIAL PRIMARY KEY,
  model      TEXT NOT NULL,
  question   TEXT,
  score      REAL,
  latency_ms INTEGER,
  created_at TIMESTAMP DEFAULT now()
);

INSERT INTO evals (model, question, score, latency_ms)
VALUES ('claude-sonnet-5', 'What is RAG?', 0.92, 840);

UPDATE evals SET score = 0.95 WHERE id = 42;
DELETE FROM evals WHERE created_at < '2025-01-01';

-- Speed up the columns you filter/join on constantly
CREATE INDEX idx_evals_model_time ON evals (model, created_at);

-- Ask the database how it will run your query
EXPLAIN ANALYZE
SELECT * FROM evals WHERE model = 'claude-sonnet-5';

-- Postgres + the pgvector extension = a vector database
CREATE EXTENSION vector;

CREATE TABLE chunks (
  id        SERIAL PRIMARY KEY,
  doc       TEXT,
  content   TEXT,
  embedding vector(768)          -- one embedding per chunk
);

-- Top-5 most similar chunks to a query embedding (cosine distance: <=>)
SELECT content, 1 - (embedding <=> $query_embedding) AS similarity
FROM chunks
ORDER BY embedding <=> $query_embedding
LIMIT 5;

Quick Reference

The cheat sheet

Reference Guide

Practice resources — verified & free

Same rules as the Resources page: every link checked live, each with a job to do.