Pinecone indexes, kept in sync.
Map a view's vector column to a Pinecone index. Inserts and updates upsert within seconds, and deletes remove the vector.
Overview
How it works.
Basalt reads the view's change feed and upserts vectors by row key into the namespace you choose. Other columns are sent as metadata for filtering. Deletes in the view become deletes in the index.
Setup
Set it up in three steps.
01
Connect the tool
Paste a Pinecone API key and name an existing serverless index. Basalt checks that dimensions match before the first upsert.
# sync product embeddings to Pinecone basalt sink add catalog --pinecone products-prod --view product_vectors --namespace live
✓ index products-prod · 1536 dims · cosine
✓ 1.2M vectors upserted · lag 1.8 s
02
Shape it with a view
Write plain SQL over the source. Basalt keeps the view current with every record and tracks lineage back to the topic, partition and offset.
CREATE MATERIALIZED VIEW revenue_by_min AS SELECT date_trunc('minute', ts) AS minute, currency, sum(amount) AS revenue, count(*) AS payments FROM payments WHERE status = 'captured' GROUP BY 1, 2;
03
Read it from anywhere
Every view is an endpoint the moment it exists. Call it over REST, connect with any Postgres client, or use the SDK from your app.
curl https://api.basalt.dev/v1/views/revenue_by_min?limit=5 \ -H "Authorization: Bearer $BASALT_TOKEN"
✓ 200 OK · 5 rows · served from eu-north-1 in 3.1 ms
Included
What you get.
01
Idempotent upserts
Vectors are keyed by row ID, so retries overwrite instead of duplicating.
02
Metadata filters
Pick columns to send as metadata for filtered queries in Pinecone.
03
Deletes follow
Rows removed from the view are deleted from the index within seconds.
04
Namespace routing
Route tenants or environments to separate namespaces from one view.
Related
