About Pipelines

Pipelines contain the processing stages needed to execute a query and index your files. These stages are pipeline components, also called nodes, that are connected in series so that the output of one node is used by the next node in the pipeline.

How Do Pipelines Work?

An icon showing nodes joined together to form a pipeline. Nodes are depicted as squares with rounded edges joined together with a dotted line.

Pipelines define how data flows through its nodes to achieve the best search results. For example, a basic pipeline can be made up of a retriever and a reader. The retriever goes through all the documents you want to use for your search and selects the most relevant ones to the query. Then, the reader uses the documents selected by the retriever and highlights the word, phrase, sentence, or paragraph that answers your query.

Nodes are like building blocks that you can mix and match or replace. They can be connected as a Directed Acyclic Graph (DAG), thus allowing for more complex workflows, such as decision nodes or having the output of multiple nodes combined.

Pipelines run on the files you add to deepset Cloud and turn them into documents. A document is a piece of text stored in the document store. Multiple documents may come from one file.

deepset Cloud currently supports two types of pipelines: question answering and information retrieval.

Indexing and Query Pipelines

To run a search in deepset Cloud, you must define two pipelines in your pipeline file:

  • A query pipeline that contains a recipe for how to execute a query.
An image depicting the query pipeline. It starts with the query "what is the capital of sudan", then there's an arrow pointing from the query to the pipeline which is a set of connected nodes that communicate with the document store. At the end, there's the answer Khartoum.
  • An indexing pipeline that defines how you want to preprocess your files before running a search on them.
An image of an indexing pipeline. At the beginning, there's an icon representing a file, the file goes into the indexing pipeline which is a set of connected nodes, and then at the end, it returns documents in the document store as the output. The document store icon is a tray of documents.

In deepset Cloud, you define the indexing and query pipelines in one file, which you later deploy to use for search.

When you deploy your pipeline, it indexes the files, turns them into documents, and stores them in the document store from where they're retrieved at the time of the search. The exact steps involved in indexing depend on the retrieval method you choose.
Your files are indexed once; they aren't indexed every time a pipeline runs. If you add a new file after you deploy your pipeline, only this file is indexed. The same is true for conversion. If you're using a Converter node in your pipeline, it converts the files only once; it doesn't convert them every time you run your search.

Example of an indexing and a query pipeline
version: '1.12.2'
name: "my_sample_pipeline"

components:    # define all the nodes that make up your pipeline:
  - name: DocumentStore
    type: DeepsetCloudDocumentStore
  - name: Retriever
    type: ElasticsearchRetriever
    params:
      document_store: DocumentStore    # params can reference other Components defined in the YAML
      top_k: 20
  - name: Reader       # custom-name for the component; helpful for visualization & debugging (coming soon)
    type: FARMReader    # Haystack class name for the Component
    params:
      model_name_or_path: deepset/roberta-base-squad2-distilled
      context_window_size: 500
      return_no_answer: true
  - name: TextFileConverter
    type: TextConverter
  - name: Preprocessor
    type: PreProcessor
    params:
      split_by: word
      split_length: 250
      language: en # Specify the language of your documents

pipelines:
# this is the query pipeline:
  - name: query    # a sample extractive-qa Pipeline
    type: Query
    nodes:
      - name: Retriever
        inputs: [Query]
      - name: Reader
        inputs: [Retriever]
# this is the indexing pipeline:
  - name: indexing
    type: Indexing
    nodes:
      - name: TextFileConverter
        inputs: [File]
      - name: Preprocessor
        inputs: [ TextFileConverter]
      - name: Retriever
        inputs: [Preprocessor]
      - name: DocumentStore
        inputs: [Retriever]

Pipeline Nodes

Nodes are the components that make up your pipeline. Choosing the right nodes for your pipeline is crucial to achieving the most relevant search results. Check the nodes available and find out about their superpowers.

Each node has different types, and each type was designed with a particular task in mind. For example, if you are looking for a retrieval method that doesn't need a neural network for indexing, you can use ElasticsearchRetriever (BM25). You can also specify parameters for your nodes to make them work exactly as you need.

When choosing a node for your pipeline, make sure:

  • It's optimal for the type of data you want to run your search on.
  • It's compatible with the data store you want to use.
  • It's already supported by deepset Cloud.

This table lists some of the nodes that you can use in your search system:

NodeAvailable types (Python classes)What's it best for?
FileTypeClassifierOnly one typeRoutes files with different extensions to appropriate file converters. Useful if you have different types of files.
TextConverterOnly one typeConverts a file to a document object.
PDFToTextConverterOnly one typeConverts a PDF file to plain text.
PreProcessorOnly one typeCleans files and splits them into documents.

Dealing with long documents can be a problem for some nodes. Long documents slow down the reader. Also, dense retrievers can only read about 500 words of a document. Use a preprocessor to get around it.
RetrieverElasticsearchRetriever (BM25)
ElasticsearchFilterOnlyRetriever
TfidfRetriever
EmbeddingRetriever
DensePassageRetriever (DPR)
Filters documents from the document store to retrieve a collection of documents relevant to the query.

When combined with a reader, it speeds up a query.

When used on its own, returns whole documents as answers.
ReaderFARMReader
TransformersReader
The core component that fetches the right answers.

Use a reader if you want your answers highlighted.
JoinDocumentsJoinDocumentsCombines the output of two or more retrievers. Useful if you want to use a keyword-based and a dense retriever in one pipeline.
Query ClassifierTransformersQueryClassifier
SkLearnQueryClassifier
Distinguishes between different types of queries and routes them to the pipeline branch that can handle them best.

Additionally, deepset Cloud currently supports the following document stores:

The Pipelines Page

All the pipelines created by your organization are listed on the Pipelines page. The pipelines listed under Deployed are the ones that you can run your search with. The pipelines under In Development are drafts you must deploy before you can use them for your search.

Pipeline Status

When you deploy a pipeline, it changes its status as follows:

  • Not indexed: The pipeline is being deployed, but the files have not yet been indexed
  • Indexing: Your files are being indexed. You can see how many files have already been indexed if you hover your mouse over the Indexing label.
  • Indexed: Your pipeline is deployed, all the files are indexed, and you can use your pipeline for search.
  • Partially indexed: At least one of the files wasn't indexed. This may be an NLP-related problem, a problem with your file, or a Node in the pipeline. You can still run a search if at least some files were indexed.
  • Failed to deploy: It's a fatal state. Your pipeline was not deployed, and your files are not indexed. For ideas on how to fix it, see Troubleshoot Pipeline Deployment.

Related Links