Turning unstructured data into embeddings a model can retrieve from

Building a RAG pipeline on Google Cloud starts with converting unstructured sources (PDFs, images, HTML) into clean text, chunking that text appropriately, and generating vector embeddings that are indexed for similarity search. Google Cloud offers both fully managed RAG solutions and composable building blocks (parsing, embedding, vector storage) depending on how much control you need over the pipeline.

1 · Learn the must-know

• 1Use Document AI to OCR and parse unstructured documents (scanned PDFs, images, forms) into structured text before chunking, since embedding models require clean text input.
• 2Split parsed text into overlapping chunks (commonly a few hundred tokens each) to stay within embedding model input limits while preserving enough context for accurate retrieval.
• 3Vertex AI's embedding models (text and multimodal) convert chunks into dense vectors; the multimodal embedding model can jointly embed text, images, and video into the same vector space.
• 4Vertex AI Vector Search (formerly Matching Engine) provides scalable, low-latency approximate nearest neighbor (ANN) retrieval over embeddings, while BigQuery and AlloyDB also support vector similarity search for SQL-integrated or smaller-scale use cases.
• 5Vertex AI Search (under Agent Builder) offers an out-of-the-box RAG solution that handles ingestion, chunking, embedding, indexing, retrieval, and grounded generation without building a custom pipeline.
• 6Store metadata (source document, page number, timestamp) alongside each embedding to support filtering and citation, and re-embed/re-index whenever source documents are updated to avoid retrieving stale content.

2 · Check your understanding