Week 2 · Field guide · RAG framework

Haystack

A RAG and search framework built around explicit, composable pipelines of typed components: clear architecture over magic, from the deepset team.

RAGopen sourcePythonpipeline-clarity

What it is

Search and RAG as an explicit pipeline.

Haystack, by deepset, models retrieval and search as a pipeline of typed components you wire together: a retriever, a ranker, a prompt builder, a generator, each with declared inputs and outputs. Nothing is hidden behind a convenience call, so the architecture is easy to read, reason about, and unit-test. The trade-off against LlamaIndex is a smaller connector ecosystem in exchange for a clearer, more explicit model of the flow.

The one job

Production search and RAG from explicit components.

It turns a search or RAG system into a graph of typed, testable parts. You can see exactly what each stage receives and returns, which makes the pipeline something you operate rather than something you hope works.

Reach for it when

You value architecture clarity and want explicit pipelines over implicit magic.
Your problem is search-heavy, where retrievers and rankers are the core.
You want typed components you can unit-test in isolation.
You need a pipeline whose data flow is visible to the whole team.

Skip it when

You need the largest connector ecosystem: LlamaIndex has far more data sources off the shelf.
A Python-only constraint is a problem: there is no first-class TypeScript path.
You already run one orchestration framework and do not want a second model of the flow.

The principle it teaches

Course principle

Explicit pipelines beat magic. Haystack mirrors the course's "make the flow visible" stance: when every stage is a named, typed component, you can test it, trace it, and reason about failures instead of guessing.

Where it fits your labs

This is a Week 2 tool, and a strong alternative to LlamaIndex for Lab 2. Reach for it when you would rather see the RAG pipeline as explicit, testable components than lean on a framework's convenience layer.

Use in

Lab 2 (hybrid RAG), as an alternative to LlamaIndex.

week 2

Pairs with

A vector database for storage and a reranker as the precision stage.