Haystack

What it is for

Haystack is an Apache-2.0 orchestration framework for RAG and agents: wire reusable components into a pipeline, or hand step-by-step control to an Agent, then run it locally or in deepset Studio. Use it when the job is a production retrieval or agent graph you own, not when you need a hosted paper-chat with no repo.

What makes it unique: the cover is the official deepset-ai/haystack GitHub file list — haystack/, examples/, Apache-2.0 — not a docs footer.

Key features

• Pipelines — a graph you wire: embedder → retriever → prompt builder → generator, with branching and loops (Introduction; Get Started). Studio’s visual builder is this graph in a GUI, not a second product.
• Agents — the LLM picks the next tool or pipeline call; docs contrast this with a fixed pipeline (Choosing Between Pipelines and Agents). A control loop, not another retriever.
• Document stores + retrieval — write embeddings, retrieve, rank. Weaviate / OpenSearch / in-memory are one retrieve stream (Introduction).
• Evaluators / groundedness — score answers against context once the graph runs (production path on the intro; deepset pricing Groundedness Observability). A score, not a generate.
• Prompt Explorer — compare prompts without rebuilding the graph (deepset.ai/pricing feature table). A prompt bench, not a pipeline run.

Playground is the same test surface as running the graph. haystack-ai install, HTTP serving, and MCP extras are install / SDK. Model names and document-store brands are SKU.

Who it's for

Best fit: teams who will write Python pipelines (or Studio) and already have a model + store.

Pricing

Verified 2026-10-22 (Asia/Shanghai) from the Apache-2.0 LICENSE and deepset.ai/pricing.

• Software — Apache-2.0. License fee $0 (haystack-ai).
• Studio — $0. 1 workspace; 1 user; 100 pipeline hours; 50 files (max 10 MB); 2 development pipelines; cloud deploy; Discord support.
• Enterprise — Custom. Unlimited workspaces / files / pipelines; production HA; dedicated team; SSO / RBAC. Contact deepset. No public seat dollar on that card.

Haystack Enterprise Starter (support hours) is a sales path on the blog — no dollar on the pricing cards we copied. Do not invent it.

Limitations / costs

• Studio is one user + 100 pipeline hours. Production HA is the custom card.
• You bring the model and the store bill on OSS.
• Framework + Studio, not a paper writer. For cited academic drafts, see Yomu.

Alternatives

• LlamaIndex — another Python data/agent framework with a hosted control plane.
• RAGFlow — open-source RAG studio with a visual canvas.
• PaperQA2 — paper-grounded QA when the corpus is PDFs, not a general pipeline kit.

Official tutorial

Official deepset, makers of Haystack channel video: Intro to Haystack Pipelines (Tuana Celik — components, pipeline.show(), first RAG).

Written path: Get Started plus Pricing. Third-party “Haystack in 30 minutes” conference rips are not embedded.

Website

https://haystack.deepset.ai

Repository

https://github.com/deepset-ai/haystack

Source

Catalog row curated for AI Tools Nav. Cover is the official deepset-ai/haystack GitHub file list (Apache-2.0). Dollars checked on deepset.ai/pricing (Studio $0; Enterprise Custom). Tutorial oEmbed author: deepset.

Last verified

2026-10-22 (Asia/Shanghai) editorial check.