Source: Vertex AI RAG Engine — Overview and Transformations

Google's managed RAG pipeline (corpus, ingestion, chunking, embedding, retrieval), with its region, CMEK and VPC-SC constraints.

Contents

Summary

Explains the RAG Engine pipeline stages and default chunking parameters, plus where it’s available.

Key claims

  • Pipeline: ingest (local files, Cloud Storage, Google Drive) → transform/chunk → embed → index into a RAG corpus → retrieve → generate with retrieved context.
  • Default chunking: chunk_size 1,024 tokens, chunk_overlap 256 tokens. Smaller chunks give more precise embeddings; larger ones more general.
  • Layout-aware parsing via the Document AI layout parser integration.
  • A RAG-Engine-managed Spanner instance can serve as the vector DB and is billed to you.
  • Access to us-central1, us-east1, us-east4 and europe-west1 requires allowlisting; the page tells users with new projects to try other regions.
  • Supports VPC-SC and CMEK; does not support data residency or Access Transparency (AXT) controls.

Feeds

Google Vertex AI

Suggest changes →