Haystack

Haystack is an Apache 2.0 licensed Python framework maintained by Berlin-based deepset. Developers use it to build LLM applications from modular components wired into pipelines. Common uses are retrieval-augmented generation, agentic RAG, semantic search, conversational assistants and multimodal workflows such as image processing and audio transcription. Pipelines can branch and loop, run synchronously or asynchronously, stream output, and be serialized for cloud-agnostic, Kubernetes-ready deployment. Agents support standardized tool calling, concurrent tool calls, lifecycle hooks for guardrails, and tracking of steps, tokens and tool calls. The companion Hayhooks project serves pipelines as HTTP endpoints or MCP servers. Haystack is model-agnostic, with integrations for OpenAI, Anthropic, Mistral and Hugging Face models and for document stores such as Weaviate, Pinecone and Elasticsearch. The framework itself is free, so the real costs are model APIs, compute and vector storage. Haystack Enterprise Starter adds pipeline templates, four hours a month of consultation with core maintainers, email support and early access to some features. It is priced by organization size and has no published rates. The separate deepset AI Platform adds a visual designer, access controls, auditability and cloud or on-premises deployment. Haystack suits engineering teams that want explicit, debuggable control over retrieval and context. Its main limitation is that it is code-first: teams without Python developers cannot use it directly, and anonymous telemetry is on by default until you opt out.