RAG with Cloud Run and Cloud SQL
May 23, 2026 ยท View on GitHub
This module deploys a "Retrieval-Augmented Generation" (RAG) system, leveraging Cloud Run, Cloud SQL and BigQuery.
A Cloud Run job periodically ingests sample movies data from BigQuery, creates embeddings and stores them in a Cloud SQL database (with pgvector). Another Cloud Run frontend application leverages the text embeddings from the Cloud SQL database and answers questions on these movies in json format.
Core Components
The deployment includes:
-
An ingestion subsystem, made of a private Cloud Run job with direct egress access to the user VPC, that reads sample data from BigQuery, leverages the Vertex Text Embeddings APIs and stores results in Cloud SQL
-
Databases:
- A BigQuery dataset, where users store their data to augment the model
- A private Cloud SQL instance where the ingestion job stores the text embeddings. By default, this is PostgreSQL with the pgvector extension enabled.
-
A frontend subsystem, made of:
- Global external application load balancer (+ Cloud Armor IP allowlist security backend policy + HTTP to HTTPS redirect + managed certificates). This is created by default.
- Internal application load balancer (+ Cloud Armor IP allowlist security backend policy + HTTP to HTTPS redirect + managed certificates + CAS + Cloud DNS private zone). This is not created by default.
- A private Cloud Run frontend with direct egress access to the user VPC, that answers user queries in json format, based on the text embeddings contained in the Cloud SQL database.
-
By default, a host project, a shared VPC, a subnet, private Google APIs routes and DNS policies. Optionally, you can use your own host project and shared VPCs.
-
A service project with all the necessary APIs, service accounts, permissions set.