Haystack (deepset)

Not the conference

This is deepset’s Python framework. The Haystack conference series is Haystack US and Haystack EU — unrelated beyond the shared name.

An open-source AI orchestration framework for building LLM applications in Python, maintained by deepset (deepset-ai), Apache-2.0. It describes itself as a framework for “context-engineered, production-ready LLM applications”, built around modular pipelines and agent workflows with explicit control over retrieval, routing, memory, and generation.

🔗 https://haystack.deepset.ai/ · https://github.com/deepset-ai/haystack

Core Abstractions

ComponentRole
PipelinesSynchronous and asynchronous composition of components
AgentsTool calling with lifecycle hooks
RetrieversRetrieval and indexing components
RoutersConditional control flow — see Query Routing
Evaluation utilitiesPipeline quality measurement

The framework supports native async execution, token-by-token streaming, and integrations across multiple model vendors (OpenAI, Mistral, Anthropic, Cohere, Hugging Face).

Why It Matters Here

Haystack’s router components are one of the more explicit treatments of Query Routing in a mainstream framework — ZeroShotTextRouter, TextClassificationRouter, ConditionalRouter and FileTypeRouter make the routing decision a declarative pipeline node rather than application code.

  • LlamaIndex · LangChain — the other two general-purpose RAG orchestration frameworks
  • RAGAS — evaluation, commonly paired with Haystack pipelines

Articles