Use Document AI layout parser with Vertex AI RAG Engine

This page introduces the Document AI layout parser and how it's used with RAG Engine.

Document AI

Document AI is a document-processing and document-understanding platform that takes unstructured data from documents and transforms that unstructured data into fields that are suitable for storage in a database. Structured data leads to data that you can understand, analyze, and consume.

Document AI is built on top of products within Vertex AI with generative AI to help you create scalable, end-to-end, cloud-based document processing applications. No specialized machine-learning expertise is required to use these products.

Document AI layout parser

The layout parser extracts content elements from the document, such as text, tables, and lists. The layout parser then creates context-aware chunks that facilitate information retrieval in generative AI and discovery applications.

When it's used for retrieval and LLM generation, the document's layout is considered during the chunking process, which improves semantic coherence and reduces noise in the content. All text in a chunk comes from the same layout entity, such as the heading, subheading, or list.

For file types used by layout detection, see Layout detection per file type.

Use the layout parser in Vertex AI RAG Engine

To use the layout parser in Vertex AI RAG Engine, you must create a corpus. To create a corpus, do the following:

1. In the Google Cloud console, go to the RAG Engine page.

   Go to RAG Engine

2. Select Create corpus.

3. In the Region field, select your region.

4. In the Corpus name field, enter your corpus name.

5. In the Description field, enter a description.

6. In the Data section, select where you want to upload your data.

7. Expand the Advanced options section.

   1. In the Chunking strategy section, the following default sizes are recommended:

      • Chunking size: 1024
      • Chunk overlap: 256

   2. In the Layout parser section, select the LLM parser option, which has the highest accuracy for documents with images or charts.

      1. From the Model field, select your model.

      2. Optional: In the Maximum parsing requests per min field, enter your maximum parsing requests.

      3. Optional: In the Custom parsing prompt field, enter your parsing prompt.

      4. Click Continue.

8. From the Configure vector store page, do the following:

   1. In the Embedding model field, select your embedding model.

   2. In the Vector database section, select your database.

9. Click Create corpus.

Limitations

The ImportRagFiles API supports the layout parser, however, the following limitations apply:

• Input the file size maximum of 20 MB for all file types.
• There is a maximum of 500 pages per PDF file.

The Document AI quotas and pricing apply.

Enable the Document AI API

You must enable the Document AI API for your project. For more information on enabling APIs, see the Service Usage documentation.

Enable the Document AI API.

Enable the API

Turn on your layout parser

To turn on Layout Parser, follow these steps:

1. Create a Layout Parser by following the instructions in Creating and managing processors.

   The processor type name is LAYOUT_PARSER_PROCESSOR.

2. Enable Layout Parser by following the instructions in Enable a processor.

Your RAG knowledge base (corpus)

If you don't have a RAG corpus, then create a RAG corpus. For example, see Create a RAG corpus example.

If you already have a RAG corpus, existing files that were imported without a layout parser won't be re-imported when you Import files using Layout Parser. If you want to use a layout parser with your files, delete the files first. For example, see Delete a RAG file example.

Importing files using layout parser

Files and folders from various sources can be imported using the layout parser.

from vertexai import rag
import vertexai

PROJECT_ID = YOUR_PROJECT_ID
corpus_name = "projects/{PROJECT_ID}/locations/us-central1/ragCorpora/{rag_corpus_id}"
paths = ["https://drive.google.com/file/123", "gs://my_bucket/my_files_dir"]  # Supports Cloud Storage and Google Drive.

# Initialize Vertex AI API once per session
vertexai.init(project=PROJECT_ID, location="us-central1")

response = rag.import_files(
    corpus_name=corpus_name,
    paths=paths,
    transformation_config=rag.TransformationConfig(
        rag.ChunkingConfig(chunk_size=1024, chunk_overlap=256)
    ),
    import_result_sink="gs://sample-existing-folder/sample_import_result_unique.ndjson",  # Optional: This must be an existing storage bucket folder, and the filename must be unique (non-existent).
    llm_parser=rag.LlmParserConfig(
      model_name="gemini-2.5-pro-preview-05-06",
      max_parsing_requests_per_min=100,
    ),  # Optional
    max_embedding_requests_per_min=900,  # Optional
)
print(f"Imported {response.imported_rag_files} files.")

POST https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/ragCorpora/RAG_CORPUS_ID/ragFiles:import

{
  "import_rag_files_config": {
    "gcs_source": {
      "uris": "GCS_URIS"
    },
    "rag_file_parsing_config": {
      "layout_parser": {
        "processor_name": "LAYOUT_PARSER_PROCESSOR_NAME"
      }
    },
    "rag_file_transformation_config": {
      "rag_file_chunking_config": {
        "fixed_length_chunking": {
          "chunk_size": CHUNK_SIZE
        }
      }
    },
  }
}

curl -X POST \
    -H "Authorization: Bearer $(gcloud auth print-access-token)" \
    -H "Content-Type: application/json; charset=utf-8" \
    -d @request.json \
    "https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/ragCorpora/RAG_CORPUS_ID/ragFiles:import"

$cred = gcloud auth print-access-token
$headers = @{ "Authorization" = "Bearer $cred" }

Invoke-WebRequest `
    -Method POST `
    -Headers $headers `
    -ContentType: "application/json; charset=utf-8" `
    -InFile request.json `
    -Uri "https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/ragCorpora/RAG_CORPUS_ID/ragFiles:import" | Select-Object -Expand Content

Retrieval query

When a user asks a question or provides a prompt, the retrieval component in RAG searches through its knowledge base to find information that is relevant to the query.

For an example of retrieving RAG files from a corpus based on a query text, see Retrieval query.

Prediction

The prediction generates a grounded response using the retrieved contexts. For an example, see Generation.

What's next

• Vector database choices in Vertex AI RAG Engine
• To learn how to import RAG files, see the Import RAG files example.