How to Install the ParadeDB Extension in PostgreSQL

What is ParadeDB and why use it?

ParadeDB is a PostgreSQL extension that layers vector similarity search and distributed full-text search on top of vanilla Postgres. It lets cloud-native apps stay on a single database instead of adding separate search services.

How do I install ParadeDB in a cloud-native Postgres cluster?

Install the extension package in your container image or VM, then run CREATE EXTENSION parade; using a superuser role. Helm charts for Kubernetes often expose a postInit SQL hook for this.

Which parameters matter during installation?

parade.indexing_parallelism controls worker threads. parade.default_distance_metric sets cosine, euclidean, or dot for vector search. Tune them with ALTER SYSTEM or per-session SET.

How do I store product embeddings?

Add a vector column to the Products table. Populate it with 1536-dimension OpenAI embeddings.

ALTER TABLE Products ADD COLUMN embedding vector(1536);
UPDATE Products
SET embedding = openai_embed(name || ' ' || description);


How do I run a similarity search?

Use the <=> operator supplied by ParadeDB.

SELECT id, name, price
FROM Products
ORDER BY embedding <=> openai_embed('wireless earbuds')
LIMIT 5;


How can Orders be joined with search results?

Because ParadeDB lives inside Postgres, you can write one query that ranks products then joins to order line items for revenue analytics.

WITH ranked AS (
SELECT id AS product_id
FROM Products
ORDER BY embedding <=> openai_embed('laptop sleeve')
LIMIT 10
)
SELECT p.name, SUM(oi.quantity) AS units_sold
FROM ranked r
JOIN OrderItems oi ON oi.product_id = r.product_id
JOIN Products p ON p.id = r.product_id
GROUP BY p.name;


When should I reindex vectors?

Reindex after bulk inserts or when recall degrades. Run SELECT parade.reindex('Products_embedding_idx'); during low-traffic windows.

Best practices for cloud-native deployments?

Use read-replicas for heavy search, pin ParadeDB worker memory with Kubernetes limits, and store embeddings in a separate tablespace backed by fast NVMe.

What are common mistakes?

Wrong distance metric: using cosine for unit-normalized vectors is fine, but dot product is faster if vectors are already normalized. Change via ALTER INDEX ... SET (parade.distance_metric='dot').

Ignoring autovacuum: large embedding columns bloat quickly. Tune autovacuum_vacuum_scale_factor lower to prevent index slowdown.

FAQs

Does ParadeDB work with Postgres 16?

Yes, ParadeDB 0.10+ supports PostgreSQL 13-16.

Can I use ParadeDB on managed services like Neon or Supabase?

Only if the provider lets you load custom extensions. For now, self-hosted Postgres or AWS RDS Custom is required.