Prepare multimodal RAG data in batches with a workflow

MOI provides workflow templates for RAG data preparation that can read documents, parse layouts, split text into chunks, generate embeddings, and build text and image indexes in a single batch run. You can apply the same processing rules repeatedly to a collection of files, making these templates suitable for continuously building a knowledge base containing text, tables, and images.

This tutorial uses 3 illustrated Landsat reports published by the U.S. Geological Survey (USGS), covering Landsat applications in Alaska, California, and Florida. Each PDF contains text, maps, satellite imagery, or thematic maps, allowing you to validate text parsing, image descriptions, and image indexing together. You will use the Document Knowledge Base Preparation (with Image Indexing) template to prepare multimodal RAG data in a batch, then verify cross-document Q&A and image retrieval. The tutorial takes approximately 30 minutes; actual processing time depends on the number of files, page counts, and model speed.

Screenshots use the Chinese interface. The downloadable sample documents and their filenames are unchanged from the Chinese tutorial.

What You Will Accomplish

  • Upload multiple PDFs containing text and images to Catalog in a batch.

  • Create a multimodal RAG data preparation workflow from a system template.

  • Configure document parsing, text indexing, and image indexing.

  • Run the workflow once to process all sample files.

  • Verify cross-document Q&A and figure retrieval results in the knowledge base.

Before You Begin

The sample package contains the following public materials:

File

Text and Visual Content

Official Source

USGS.Alaska-and-Landsat.pdf

Text, a near-surface permafrost distribution map, satellite imagery of flooding, and false-color wildfire imagery

Alaska and Landsat

USGS.California-and-Landsat.pdf

Text, false-color wildfire imagery, and a vegetation ecoregion map

California and Landsat

USGS.Florida-and-Landsat.pdf

Text, a map showing fire years, mangrove photographs, and satellite imagery of Miami Beach

Florida and Landsat

These files are provided for convenience when following the tutorial. For your own use, you can replace them with your PDF, Word, PowerPoint, web page, spreadsheet, or image materials.

Steps

1. Upload the Sample Files to Catalog in a Batch

  1. Sign in to MOI and select Import Tasks under Data Connect in the left navigation.

  2. Click New Import Task and select Unstructured.

  3. Select Local Upload as the data source, then select the extracted USGS.Alaska-and-Landsat.pdf, USGS.California-and-Landsat.pdf, and USGS.Florida-and-Landsat.pdf files together.

  4. Select the current workspace’s Catalog as the target and save the files to:

    <Current workspace Catalog> / tutorial_data / multimodal_rag / source

    Here, tutorial_data is the catalog, multimodal_rag is the database, and source is the data volume. If they do not exist, create each one while selecting the target location.

  5. Keep the remaining settings at their defaults and confirm that all 3 PDFs appear in the import file list.

    Configure Local Upload and select three USGS Landsat PDFs containing text and images at once

  6. Click Create and Start Import.

  7. Wait until the import task status changes to Completed and confirm that all 3 PDFs are available at the following location:

    The import task for the three USGS Landsat PDFs containing text and images has a Completed status 
    Resource Center > Catalog > tutorial_data > multimodal_rag > source

2. Create a Multimodal RAG Workflow

  1. Select Workflow under Data Processing in the left navigation.

  2. Click Create Workflow.

  3. Find the Document Knowledge Base Preparation (with Image Indexing) template and click Use Template.

    Select the Document Knowledge Base Preparation (with Image Indexing) template when creating a workflow

  4. Enter landsat_visual_rag as the workflow name.

The template already includes nodes for reading files, parsing documents, splitting text, building text and image indexes, saving parsing results, and registering data lineage. You do not need to add these nodes individually.

3. Configure the Workflow

Select the Source Files for Batch Processing

  1. Open the workflow’s run configuration and select Catalog Data Source on the left.

  2. Under Source Files, select the tutorial_data / multimodal_rag / source data volume.

  3. Set File Scope to By File and select USGS.Alaska-and-Landsat.pdf, USGS.California-and-Landsat.pdf, and USGS.Florida-and-Landsat.pdf.

    Select three PDFs containing text and images from the source data volume in Catalog

Selecting all files at once lets the workflow process them in a batch during the same run.

Create a Knowledge Base

  1. Select Build Knowledge Base Index on the left.

  2. Enter landsat_environment_knowledge under Knowledge Base and click Create Knowledge Base.

  3. Keep the default text Embedding model, turn on Enable Image Indexing, and keep the default image Embedding model.

  4. Keep the three-level index and remaining parameters at their defaults.

    Create a multimodal RAG knowledge base and enable image indexing

Save to Catalog

  1. Select Save to Catalog on the left.

  2. Set Output Location to tutorial_data / multimodal_rag / source.

  3. Click Confirm to save the run configuration.

    Save workflow processing results to the source data volume in Catalog

Keep the template’s default configuration for the remaining steps, including Document Parsing, Split Documents by Length, Engineering Drawing Parsing, Build Drawing Knowledge Base Index, and Write Document Files. You do not need to modify each step individually.

4. Save and Run the Workflow in a Batch

  1. Return to the workflow canvas and click Save and Enable.

  2. Click Run on the workflow page.

  3. Confirm that both the source and output locations are source and that all 3 PDFs are selected.

  4. After the run starts, click Jobs on the right side of the workflow page.

  5. Find this run in the job list and wait until the job statuses for all 3 files change to Completed.

    The job list shows a Completed run status for landsat_visual_rag

Only one run is needed. The workflow processes all input files using the same parsing and indexing rules.

5. Start a Conversation in the Knowledge Base

  1. Select Knowledge Base in the left navigation and open landsat_environment_knowledge.

    Select landsat_environment_knowledge in the knowledge base list and click Chat

  2. Click Chat on the knowledge base card and send a cross-document question:

    What environmental issues are monitored using Landsat in Alaska, California, and Florida? Summarize them by region.
    The knowledge base answers a cross-document question about Landsat environmental monitoring and provides references

  3. Click the sources in the answer to view the cited materials and the original passages.

  4. Send another question about the figures:

    Find the false-color satellite image of California's August Complex wildfire and explain what yellow and dark brown represent in the image.
    The knowledge base answers a question about Landsat false-color satellite imagery

  5. Click the page or image sources in the answer to view the captions and source information.

Tutorial Complete

You have used a system template to prepare multimodal RAG data in a batch. All 3 public documents were parsed, chunked, and indexed for both text and images by the same workflow. The parsing results are saved to Catalog and can be used in the knowledge base for cross-document Q&A and figure retrieval.

Last updated on