haystack

Haystack Expert Skill

Haystack (by deepset) is a production-oriented framework for building search and NLP pipelines. Its core abstraction is the Pipeline — a directed acyclic graph of typed components with explicit connections. Unlike LangChain's LCEL (pipe operator) or LlamaIndex's query engines, Haystack pipelines are declared upfront with add_component and connect, giving validated, debuggable DAGs.

Core Paradigm

from haystack import Pipeline
from haystack.components.embedders import SentenceTransformersTextEmbedder
from haystack.components.retrievers.in_memory import InMemoryEmbeddingRetriever
from haystack.components.builders import PromptBuilder
from haystack.components.generators import OpenAIGenerator
from haystack.document_stores.in_memory import InMemoryDocumentStore

# Build a pipeline
document_store = InMemoryDocumentStore()
pipeline = Pipeline()
pipeline.add_component("embedder", SentenceTransformersTextEmbedder())
pipeline.add_component("retriever", InMemoryEmbeddingRetriever(document_store=document_store))
pipeline.add_component("prompt_builder", PromptBuilder(template="Answer using: {{documents}}\n\nQuestion: {{question}}"))
pipeline.add_component("generator", OpenAIGenerator())

# Connect components
pipeline.connect("embedder.embedding", "retriever.query_embedding")
pipeline.connect("retriever.documents", "prompt_builder.documents")
pipeline.connect("prompt_builder", "generator")

# Run
result = pipeline.run({"embedder": {"text": "What is Haystack?"}, "prompt_builder": {"question": "What is Haystack?"}})

Core Principles

1. Pipelines are validated DAGs. add_component + connect. Pipeline validation catches errors BEFORE execution — leverage this during development.
2. Components are typed. Each component has input/output slots. Connections must match types. This prevents runtime errors.
3. PromptBuilder uses Jinja2. Templates are Jinja2 strings, not f-strings. {{documents}}, {{query}}, {{question}} are variable placeholders.
4. Indexing and query are separate pipelines. One pipeline loads/cleans/embeds/writes documents. Another retrieves/generates answers. They share the DocumentStore.
5. Evaluation is a pipeline too. Add evaluator components to measure faithfulness, relevancy, or custom metrics.