====== Haystack ======
**Haystack** is an open-source AI orchestration framework by deepset for building customizable, production-ready LLM applications through [[modular|modular]] pipelines. With **24.6K [[github|GitHub]] stars** and active development since 2019, it is one of the longest-standing AI frameworks, evolving from a question-answering system to a comprehensive pipeline-based orchestration platform.

{{tag>framework python pipelines rag production orchestration deepset}}

===== Overview =====
Haystack was created by deepset (Berlin, Germany) in 2019 as an open-source question-answering framework focused on extractive QA over documents. Over six years, it has evolved through multiple paradigm shifts — from neural search with dense retrievers (2020-2021) to RAG pipelines (2022-2023) to the fully redesigned Haystack 2.0 with component-based architecture and agent workflows (2023-2024). The framework emphasizes production-readiness with built-in observability, async execution, and Kubernetes integration.(([[https://github.com/deepset-ai/haystack|GitHub Repository]]))(([[https://haystack.deepset.ai|Documentation]]))(([[https://www.deepset.ai|deepset (Maintainer]]))

===== Key Features =====
  * **[[modular|Modular]] Pipelines** — Directed acyclic graphs (DAGs) of components for RAG, QA, [[semantic_search|semantic search]], and more
  * **200+ Components** — Document stores, [[embeddings|embeddings]], LLMs, retrievers, readers, converters, and evaluators
  * **Production-Ready** — OpenTelemetry tracing, retries, caching, error handling, Kubernetes/Docker deployment
  * **Agent Workflows** — Tool-using agents with reasoning loops and conditional logic
  * **[[hybrid_search|Hybrid Search]]** — Combine sparse (BM25) and dense (DPR) retrieval with [[reranking|reranking]]
  * **Component-Based** — Stateless, composable components with Pydantic-validated I/O schemas
  * **Evaluation Metrics** — Built-in SquadMetric, ExactMatchMetric, and custom evaluators
  * **Document Stores** — Pluggable backends: Elasticsearch, OpenSearch, [[pinecone|Pinecone]], [[weaviate|Weaviate]], Chroma, In-Memory

===== Architecture =====
Haystack's pipeline-centric architecture:

<mermaid>
graph LR
    A[Preprocessor: Document Converter] --> B[Retriever: BM25 / Dense / Hybrid]
    B --> C[Reader: LLM or Extractive]
    C --> D[Postprocessor / Ranker]
</mermaid>

Infrastructure Layer:

<mermaid>
graph TD
    A[Document Store: Elasticsearch / [[pinecone|Pinecone]]]
    B[Embedding Models: Sentence Transformers / [[openai|OpenAI]]]
    C[LLMs: [[openai|OpenAI]] / [[anthropic|Anthropic]] / HuggingFace / Local]
    A, B, C
</mermaid>

===== Code Example =====
Building a RAG pipeline with Haystack 2.x:

<code python>
from haystack import Pipeline
from haystack.components.generators import OpenAIGenerator
from haystack.components.builders.prompt_builder import PromptBuilder
from haystack.components.retrievers.in_memory import InMemoryBM25Retriever
from haystack.document_stores.in_memory import InMemoryDocumentStore
from haystack.dataclasses import Document

# Set up document store with sample data
doc_store = InMemoryDocumentStore()
doc_store.write_documents([
    Document(content="Haystack is an AI orchestration framework by deepset."),
    Document(content="It supports [[modular|modular]] pipelines for RAG and search."),
    Document(content="Haystack 2.0 introduced component-based architecture."),
])

# Build RAG pipeline
template = """
Given these documents, answer the question.
Documents: {% for doc in documents %}{{ doc.content }}{% endfor %}
Question: {{ question }}
"""

rag_pipeline = Pipeline()
rag_pipeline.add_component("retriever", InMemoryBM25Retriever(document_store=doc_store))
rag_pipeline.add_component("prompt", PromptBuilder(template=template))
rag_pipeline.add_component("llm", OpenAIGenerator(model="gpt-4o"))
rag_pipeline.connect("retriever", "prompt.documents")
rag_pipeline.connect("prompt", "llm")

result = rag_pipeline.run({
    "retriever": {"query": "What is Haystack?"},
    "prompt": {"question": "What is Haystack?"}
})
print(result["llm"]["replies"][0])
</code>

===== Haystack vs LangChain =====
^ Aspect ^ Haystack ^ [[langchain|LangChain]] ^
| **Core Paradigm** | Pipeline DAGs with visual UI | Chains/Agents with LCEL |
| **Modularity** | Strong typing, 200+ components | Flexible, vast integrations |
| **Production** | Built-in observability, K8s, async | Requires [[langsmith|LangSmith]]/LangServe |
| **RAG Focus** | Optimized for search/retrieval | General-purpose agents |
| **History** | Since 2019 (6+ years) | Since 2022 |
| **Stars** | 24.6K (steady growth) | 131K (larger community) |

===== Timeline =====
  * **2019** — Initial release as extractive QA framework
  * **2020-2021** — Dense retrievers (DPR), Transformers integration, v1.0 stable
  * **2022** — Pivot to RAG amid LLM rise, [[openai|OpenAI]] GPT support
  * **2023** — Haystack 2.0 overhaul with component-based design, agents, multi-[[modal|modal]]
  * **2024-2025** — Enterprise features, advanced orchestration, self-improving pipelines
  * **2026** — Milestone #2: context-engineered LLM apps, improved agent workflows

===== See Also =====
  * [[ai_orchestration_layers|AI Orchestration Layers]]
  * [[pipecat|Pipecat]]
  * [[tool_use_orchestration|Tool Use and Orchestration]]
  * [[swarm_openai|OpenAI Swarm]]
  * [[agentic_orchestration_platforms|Agentic Orchestration Platforms Comparison]]

===== References =====