How to Use ParadeDB on Azure in PostgreSQL

What is ParadeDB on Azure?

ParadeDB is a PostgreSQL extension that adds vector search, hybrid search, and full-text search. Deploying it on Azure Database for PostgreSQL gives SaaS teams Google-like relevance without moving data to a separate vector store.

How do I enable ParadeDB in Azure Database for PostgreSQL?

Create a Flexible Server with the "Extensions" feature turned on. Connect with psql or Galaxy and run CREATE EXTENSION IF NOT EXISTS paradedb;. The command loads C functions and SQL operators needed for vector search.

How do I store product embeddings?

Create a vector column. Populate it from your ML pipeline or OpenAI API.

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

How do I run a similarity search?

Use paradedb.ivfflat_ops or paradedb.hnsw_ops for index-accelerated ANN queries.

SELECT id, name, price
FROM Products
ORDER BY embedding <-> paradedb.embed('iPhone 14 128GB')
LIMIT 5;

What syntax activates vector search?

Create an index for millisecond-level latency.

CREATE INDEX products_embedding_hnsw
ON Products USING hnsw (embedding paradedb.hnsw_ops)
WITH (m = 16, ef_construction = 200);

How do I mix keyword and vector filters?

Combine standard WHERE clauses with ANN ORDER BY.

SELECT id, name, price
FROM Products
WHERE stock > 0
ORDER BY embedding <#> paradedb.embed('budget wireless headphones')
LIMIT 10;

Best practices for performance?

Keep vectors small (≤1536 dims), use HNSW for top-k <100, and schedule VACUUM ANALYZE. On Azure, pick a server with enough vCPUs to match index build threads.

What mistakes should I avoid?

See the "Common Mistakes" section below for quick fixes.

Common Mistakes

Missing extension on replica?

Read replicas need CREATE EXTENSION paradedb; too. Without it, Logical Replication fails.

Index on wrong operator class?

Using vector_ops instead of paradedb.hnsw_ops disables ANN acceleration. Recreate with the correct class.

FAQs

Does ParadeDB work with Azure Hyperscale?

Yes, but you must load the extension on every worker node.

Can I store embeddings in a separate table?

Yes. Use a foreign key to the main table and join during searches.

How large can a vector be?

ParadeDB supports up to 16,000 dimensions, but latency grows. Stay under 3k for OLTP workloads.