Skip to main content

Generate Batched RAG Outputs from Datasets

This page explains how to generate batched RAG (Retrieval-Augmented Generation) outputs from datasets using Transformer Lab.

Prerequisites​

Before using the RAG Batched Outputs Generator, you need:

  1. A dataset containing input questions
  2. The RAG Plugin installed in Transformer Lab
  3. A running foundation model (you can enable this from the Foundation tab)

Step 1: Download the RAG Batched Outputs Generator Plugin​

  • Go to the Plugins Tab
  • Use the filter by type generator to narrow down the list
  • Find and download the RAG Batched Outputs Generator plugin
Plugin Download Gif

Step 2: Create a Generation Task​

  • Navigate to the Generate Tab
  • Click on Create Task
  • From the drop-down menu, select RAG Batched Outputs Generator
  • A pop-up window will appear for configuring your generation task

Step 2.1: Configure Your Task​

Name Your Generation Task​

  • In the first tab of the pop-up window, enter a name for your generation task

Plugin Configuration​

  • Move to the next tab labeled Plugin Config
  • Configure the following settings:
    • Input Field: Specify the column name in your dataset containing the input questions
    • RAG Parameters:
      • Set retrieval options (number of chunks, similarity threshold, etc.)
      • Configure generation parameters
    • Reranker Options:
      • Toggle Use Reranker to enable/disable reranking of retrieved documents
      • If enabled, enter the name of the reranker model from HuggingFace or the path to the reranker

Selecting Your Dataset​

  • Navigate to the Dataset tab
  • Select the dataset containing your input questions
Task Configuration Gif

Step 3: Run the Task​

  • Once you have configured your task, click on the Create button
  • Click on the Queue button to start the generation process
  • The task will begin processing each input in your dataset through the RAG pipeline
Task Configuration Gif

Step 4: View the Generated Output​

  • When the task completes, a new dataset will be created in the Datasets tab
  • This dataset will contain:
    • Your original data columns
    • The RAG-generated outputs
    • Retrieved context used for generation
    • If your input dataset was created with the "RAG QA Dataset Generator for RAG Evaluation" plugin, an expected_output column will also be preserved for evaluation purposes
Results Viewing Gif

Use Cases​

This plugin is particularly useful for:

  • Batch processing queries through your RAG system
  • Testing RAG performance on large datasets
  • Preparing data for RAG system evaluation
  • Comparing different RAG configurations