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
| Component | Role |
|---|---|
| Pipelines | Synchronous and asynchronous composition of components |
| Agents | Tool calling with lifecycle hooks |
| Retrievers | Retrieval and indexing components |
| Routers | Conditional control flow — see Query Routing |
| Evaluation utilities | Pipeline 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.
Related Tools
- LlamaIndex · LangChain — the other two general-purpose RAG orchestration frameworks
- RAGAS — evaluation, commonly paired with Haystack pipelines
Related Concepts
Articles
- Routing in RAG Driven Applications — covers Haystack’s four router components