# Prompt-to-PR: Add PostgreSQL Full-Text Search

> SOP for adding native PostgreSQL full-text search with tsvector, GIN index, ts_rank, and a Next.js search API — no third-party search service needed.

**Type:** Playbook  
**Tools:** Cursor, Claude Code, Codex, Windsurf  
**Stack:** Next.js, PostgreSQL, TypeScript  
**Difficulty:** hard  
**Updated:** 2026-06-08

---

PostgreSQL's built-in full-text search handles most product search needs without Elasticsearch or Algolia. This playbook adds a `tsvector` column, a GIN index, and a ranked search query behind a Next.js API route.

## 1. Requirement

Add full-text search over a `posts` table (columns: `title`, `body`, `tags`). Search should return results ranked by relevance with snippet highlights. The implementation must use a generated `tsvector` column (not computed at query time) so the GIN index is used.

## 2. First Prompt

```txt title="First Prompt"
Add native PostgreSQL full-text search to the posts table in this project.

Database: PostgreSQL. ORM/query builder: [Drizzle / postgres.js — use whichever
is already in src/db/].

Step 1 — migration:
  Create a migration file (or Drizzle schema change) that:
  a. Adds a generated column:
       search_vector tsvector GENERATED ALWAYS AS (
         setweight(to_tsvector('english', coalesce(title, '')), 'A') ||
         setweight(to_tsvector('english', coalesce(body, '')), 'B') ||
         setweight(to_tsvector('english', coalesce(tags, '')), 'C')
       ) STORED;
  b. Creates a GIN index on search_vector:
       CREATE INDEX posts_search_idx ON posts USING gin(search_vector);

Step 2 — query function:
  Create `src/lib/queries/search.ts` exporting `searchPosts(q: string, limit = 10)`.
  The query must:
  - Convert the user query to a tsquery: `websearch_to_tsquery('english', $1)`.
  - Filter: `search_vector @@ query`.
  - Rank: `ts_rank(search_vector, query) DESC`.
  - Return: id, title, ts_headline('english', body, query,
      'MaxWords=30, MinWords=15, ShortWord=3, HighlightAll=false') AS snippet.
  - Use parameterized query (no string interpolation of the user input).

Step 3 — API route:
  Create `src/app/api/search/route.ts` (GET). Read `q` from URL search params.
  Return 400 if q is empty or shorter than 2 chars. Return JSON array of results.
  Add cache-control: public, max-age=60.

Step 4 — do not touch any UI components.
```

## 3. Expected File Changes

```txt
src/db/migrations/<timestamp>_add_search_vector.sql   (new — or Drizzle migration)
src/db/schema.ts                                       (search_vector column if Drizzle)
src/lib/queries/search.ts                              (new — searchPosts function)
src/app/api/search/route.ts                            (new — GET endpoint)
```

## 4. Review Checklist

- `GENERATED ALWAYS AS ... STORED` — column is stored (not virtual), so the GIN index can be used. Confirm `STORED` keyword is present.
- `websearch_to_tsquery` is used (not `to_tsquery`) — handles multi-word and phrase queries from untrusted input without syntax errors.
- The user-supplied query is passed as a SQL parameter, never interpolated.
- `ts_rank` orders results — not alphabetical or insertion order.
- `ts_headline` length is capped (MaxWords=30) to avoid returning huge snippets.
- GIN index name is explicit — easier to drop/recreate if needed.
- API route validates minimum query length (2 chars) to avoid full-table-scan tsqueries like `'a':*`.
- `Cache-Control: public, max-age=60` is set on the response — search results can be short-lived cached.

## 5. Test Commands

```bash
# Run the migration
npx drizzle-kit migrate
# or: psql $DATABASE_URL < migration.sql

# Verify the generated column and index exist
psql $DATABASE_URL -c "\d posts" | grep search_vector
psql $DATABASE_URL -c "\di posts_search_idx"

# Query performance — confirm index scan, not seq scan
psql $DATABASE_URL -c "
  EXPLAIN ANALYZE
  SELECT id FROM posts
  WHERE search_vector @@ websearch_to_tsquery('english', 'your test query')
  LIMIT 10;
" | grep "Index Scan"

# Test the API endpoint
bun dev &
curl "http://localhost:3000/api/search?q=your+test+query" | jq .

# Confirm empty query returns 400
curl -o /dev/null -w "%{http_code}" "http://localhost:3000/api/search?q="
# Expect: 400
```

## 6. Common Failures

- **Seq scan instead of index scan** — `VIRTUAL` column (not `STORED`). Only `STORED` generated columns can be indexed in PostgreSQL. Confirm the migration uses `STORED`.
- **`to_tsquery` throws on multi-word input** — `to_tsquery('english', 'quick brown')` is a syntax error. Always use `websearch_to_tsquery` for user-supplied strings.
- **`ts_headline` returns entire body** — `HighlightAll=true` accidentally set, or `MaxWords` not set. Check the options string.
- **GIN index not used for `LIKE` queries** — ensure the WHERE clause uses `@@` operator, not `LIKE` or `ILIKE`. Full-text search and `LIKE` are separate.
- **Migration fails on existing data** — the `GENERATED` column populates on creation; on large tables this can be slow. Run in a transaction with a progress check.

## 7. Fix Prompt

```txt title="Fix Prompt"
EXPLAIN ANALYZE shows a sequential scan instead of an index scan on the
posts table. The search_vector column was added as VIRTUAL (or without
the STORED keyword) so PostgreSQL cannot create a GIN index on it.

Fix the migration:
  ALTER TABLE posts DROP COLUMN search_vector;
  ALTER TABLE posts ADD COLUMN search_vector tsvector
    GENERATED ALWAYS AS (
      setweight(to_tsvector('english', coalesce(title, '')), 'A') ||
      setweight(to_tsvector('english', coalesce(body, '')), 'B')
    ) STORED;
  CREATE INDEX posts_search_idx ON posts USING gin(search_vector);

Confirm the word STORED appears in the column definition.
```

## 8. PR Description

```md title="PR description"
## Feature: PostgreSQL native full-text search on posts

- Generated `tsvector` column (`STORED`) with weighted fields:
  title (A), body (B), tags (C)
- GIN index `posts_search_idx` — index scans confirmed via `EXPLAIN ANALYZE`
- `searchPosts(q, limit)` uses `websearch_to_tsquery` (safe for user input)
  and `ts_rank` for relevance ordering
- `ts_headline` snippets (max 30 words) with matched terms highlighted
- GET `/api/search?q=` — validates query length, returns ranked JSON,
  `Cache-Control: public, max-age=60`

No third-party search dependency added.
```