To set up a pipeline in Haystack, you start by creating a Pipeline object and adding components (called “nodes”) to it. Haystack pipelines define a sequence of steps for processing documents or answering questions, such as retrieving documents and analyzing them. Each node in the pipeline performs a specific task—like text extraction, retrieval, or answer generation—and you connect these nodes to control how data flows between them. For example, a basic question-answering pipeline might include a retriever node to find relevant documents and a reader node to extract answers from those documents. You define the pipeline structure by adding nodes and specifying their connections using the pipeline’s add_node() method.
Next, configure each component in the pipeline with the necessary parameters. For instance, a retriever node might use Elasticsearch or a dense vector search method like DPR (Dense Passage Retrieval), requiring you to initialize it with a document store and a model. A reader node could use a Hugging Face Transformers model (e.g., bert-base-uncased) to extract answers from retrieved text. Each component’s settings depend on your use case: you might adjust the top_k parameter to control how many documents the retriever returns or set confidence thresholds for the reader. Initialization typically involves passing these parameters when creating the component, such as Retriever(document_store=es_store, top_k=5).
Finally, customize the pipeline for advanced workflows. Haystack allows you to add preprocessing steps (e.g., file converters, text splitters) or postprocessing logic (e.g., answer filtering). For example, you could add a TextConverter node to extract text from PDFs before sending it to the retriever, or a JoinAnswers node to combine results from multiple reader models. You can also create custom nodes by subclassing BaseComponent to handle unique tasks. Pipelines are executed by calling run() with inputs like a query string or file paths. Testing and iterating on the pipeline’s structure—such as adjusting the order of nodes or swapping components—helps optimize performance for tasks like search, summarization, or QA.