Capability · Framework — rag

Haystack

Haystack (now at 2.x) is deepset's modular framework for LLM applications. You compose pipelines from typed components — retrievers, generators, rankers, converters, routers — and run them as graphs. Haystack is RAG-first but supports agents, multimodal, and tool use. It ships integrations for OpenAI, Anthropic, Cohere, Hugging Face, and major vector stores, plus deepset Cloud for managed deployments.

Framework facts

Category: rag
Language: Python
License: Apache 2.0
Repository: https://github.com/deepset-ai/haystack

Install

pip install haystack-ai

Quickstart

from haystack import Pipeline
from haystack.components.generators import OpenAIGenerator
from haystack.components.builders import PromptBuilder

pipe = Pipeline()
pipe.add_component('prompt', PromptBuilder(template='Answer: {{question}}'))
pipe.add_component('llm', OpenAIGenerator(model='gpt-4o-mini'))
pipe.connect('prompt', 'llm')
result = pipe.run({'prompt': {'question': 'What is RAG?'}})

Alternatives

LlamaIndex — more RAG-opinionated
LangChain — broader ecosystem
R2R — RAG server out of the box
txtai — embeddings-first alternative

Frequently asked questions

Haystack or LangChain for RAG?

Haystack's typed pipeline model is strict and easier to test and deploy — a good choice for enterprise RAG. LangChain has a larger ecosystem and more experimental features. Many teams pick Haystack when they need predictable production pipelines.

What's the difference between Haystack 1.x and 2.x?

Haystack 2.x (released 2024) is a ground-up rewrite with a cleaner component model, typed connections, serialisation, and better agent support. 1.x is in maintenance mode — new apps should start on 2.x.

Sources

Haystack — docs — accessed 2026-04-20
Haystack — GitHub — accessed 2026-04-20