Haystack Documentation
by deepset
A production-minded framework for composable RAG pipelines and agents you can actually hook into.
Overview
Haystack is deepset's open-source framework for production LLM applications — Apache-2.0, roughly 26k GitHub stars — and its documentation is organised around the framework's own vocabulary rather than around use cases: Concepts, Document Stores, Pipeline Components, Tools, Memory Stores, Optimization and Evaluation, and Development and Logging. The central idea is explicit composition: you wire retrievers, rankers, prompt builders and generators into a Pipeline you can inspect, branch and serialise, instead of handing control flow to a framework-owned chain. Current docs are version 3.0, released in July 2026, which is a substantial break from 2.x and worth reading with the release notes open. In 3.0 the Agent gained a hooks system (before_llm, before_tool, after_tool, on_exit, before_run, after_run) so you can inject human-in-the-loop approval, offload large tool results, or enforce a budget without patching internals; agents expose step_count, token_usage and tool_call_counts for runtime introspection; SkillToolset adds progressive disclosure so a model sees only skill names and descriptions until it needs the full instructions; and AsyncPipeline merged into a single Pipeline exposing both run and run_async. Legacy generators were removed in favour of chat interfaces, about thirty components moved out to haystack-core-integrations, and components now create external resources in warm_up() rather than __init__(). Model support spans OpenAI, Anthropic, Google, Mistral, Cohere, Hugging Face, Azure OpenAI, AWS Bedrock and local models. The 2.x and 1.x documentation remains archived and separately browsable.
At a Glance
- Topic
- Frameworks
- Level
- Intermediate
- Format
- Documentation
- Cost
- Free
- Duration
- Self-paced reference, continuously updated — currently documenting Haystack 3.0
- Provider
- deepset
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Compose retrievers, rankers, prompt builders and generators into explicit pipelines
- ✓Choose and configure a document store for your retrieval workload
- ✓Build agents and attach tools with typed inputs and outputs
- ✓Intercept agent behaviour with before_llm, before_tool and on_exit hooks
- ✓Track step_count, token_usage and tool_call_counts to enforce runtime budgets
- ✓Evaluate and optimise a pipeline instead of eyeballing its answers
- ✓Migrate a Haystack 2.x pipeline onto the unified 3.0 Pipeline API
Highlights
- •Explicit, inspectable pipelines rather than an opaque chain abstraction
- •Hooks make human-in-the-loop and budget enforcement first-class, not a workaround
- •Apache-2.0 with ~26k GitHub stars, backed by deepset as a commercial company
- •Archived 1.x and 2.x docs stay online — real help when 3.0 breaks your code
- •Model-agnostic: OpenAI, Anthropic, Google, Mistral, Cohere, Bedrock or local
Who It’s For
Best For
- ✓Teams putting a RAG or search system into production rather than a demo
- ✓Engineers who want to own the control flow of their own pipeline
- ✓Python shops evaluating Haystack against LangChain and LlamaIndex
- ✓Developers upgrading an existing Haystack 2.x codebase to 3.0
Prerequisites
- •Intermediate Python including async — pipelines expose run_async
- •Understanding of embeddings and vector retrieval basics
- •An LLM provider API key, or a local model, to run anything end to end
FAQ
What is Haystack Documentation?
The official documentation for Haystack, deepset's Apache-2.0 orchestration framework for RAG, agents and multimodal search. It documents the component-and-pipeline model, document stores, tools, memory stores, evaluation and logging, and now covers Haystack 3.0's hook-driven agents. The 2.x and 1.x docs stay archived, which matters because 3.0 broke a lot.
Is Haystack Documentation free?
Haystack Documentation is free to access.
What level is Haystack Documentation for?
Haystack Documentation is aimed at a intermediate audience. Recommended background: Intermediate Python including async — pipelines expose run_async, Understanding of embeddings and vector retrieval basics, An LLM provider API key, or a local model, to run anything end to end.
How long does Haystack Documentation take?
Expect roughly Self-paced reference, continuously updated — currently documenting Haystack 3.0. Most learners work through it at their own pace.
What will I learn from Haystack Documentation?
You'll learn: Compose retrievers, rankers, prompt builders and generators into explicit pipelines; Choose and configure a document store for your retrieval workload; Build agents and attach tools with typed inputs and outputs; Intercept agent behaviour with before_llm, before_tool and on_exit hooks; Track step_count, token_usage and tool_call_counts to enforce runtime budgets; Evaluate and optimise a pipeline instead of eyeballing its answers; Migrate a Haystack 2.x pipeline onto the unified 3.0 Pipeline API.
Topics
Sources
This page was written from 3 sources, 2 on domains other than docs.haystack.deepset.ai.