Indexing Pipeline for Preprocessing

The indexing pipeline converts your files to documents, preprocesses them, and finally stores them in the DeepsetCloudDocumentStore. The query pipeline then uses the documents from the DocumentStore for search.

deepset Cloud offers preprocessing nodes that you can add to your indexing pipeline. This way, when the pipeline runs, your files are automatically converted, split, and cleaned. Pipeline templates available in deepset Cloud include indexing pipelines that preprocess TXT, PDF, MD, DOCX, and PPTX files out of the box. For other formats, you may need to do some preprocessing outside of deepset Cloud.

To learn more about files and documents, see Basic Concepts.

How to Prepare Your Files

Here's an outline of how to plan file preprocessing:

Your files determine which nodes to use in the indexing pipeline:

• If all your files are of one type, use a file converter appropriate for handling this type as the first node in your indexing pipeline. For supported converters, see Converters,
• If you have multiple file types, use FileTypeClassifier as the first node in your indexing pipeline, and a file converter as the second node. FileTypeClassifier classifies your files based on their extension and sends them to the converter that can best handle them.

The converter's task is to convert your files into documents. However, the documents you obtain this way may not be of the optimal length for the retriever you want to use and may still need to be cleaned up. PreProcessor is the node that handles the cleaning and splitting of documents. It removes headers and footers, which is useful for not breaking up the flow of sentences across pages, it deletes empty lines, and splits your documents into smaller ones.

Smaller documents speed up your pipeline. They're also optimal for dense retrievers, which often can't handle longer text passages. For example, DensePassageRetriever was trained on documents 100-words long. That's the setting we recommend for dense retrievers. Sparse retrievers can work on slightly longer documents of around 200-300 words.

Use these suggestions as a starting point for your indexing pipeline. You may need to experiment with your settings to reach the optimal values for your use case.

For examples of indexing pipelines, see Sample Pipelines.

Pipeline Nodes for Preprocessing

There are a number of nodes that you can use in your indexing pipeline to preprocess your files. Have a look at this table to help you choose the right nodes: