Haystack is an open-source Python framework for building production-ready AI agents, retrieval-augmented generation (RAG) pipelines, and context-engineered systems. It provides modular, composable building blocks that let developers orchestrate every step of their AI workflows—from retrieval and reasoning to memory and tool use—while integrating freely with OpenAI, Anthropic, Mistral, and other providers without vendor lock-in. Teams at Lufthansa, Accenture, NVIDIA, and Airbus use Haystack to move from prototype to production with unified tooling, full visibility for debugging, and enterprise-scale reliability.
Haystack is an open-source Python framework built by deepset for constructing production-ready AI agents, retrieval-augmented generation (RAG) pipelines, and context-engineered systems. It provides modular, composable building blocks that enable developers to orchestrate every step of their AI workflows—from retrieval and reasoning to memory and tool use—while maintaining full visibility for debugging and optimization. The framework integrates seamlessly with leading LLM providers including OpenAI, Anthropic, Mistral, and Hugging Face, as well as vector databases like Weaviate, Pinecone, and Elasticsearch, eliminating vendor lock-in through its open architecture. Haystack accelerates the path from prototype to production by offering unified tooling for building, testing, and deploying AI applications, with pipelines that are serializable, cloud-agnostic, and Kubernetes-ready for enterprise-scale operations. Use cases span content generation with Jinja-2 templating, advanced RAG with hybrid retrieval and self-correction loops, multi-step AI agents with standardized tool calling, multimodal AI applications, and conversational systems. Teams at Lufthansa, Accenture, NVIDIA, and Airbus rely on Haystack for its transparency, composability, and production reliability. The framework is available as free open-source software via pip install, with optional enterprise support and the deepset Studio platform offering visual pipeline design, data workflows, and secure deployment options for organizations requiring additional services.
deepset explicitly retains broad rights to use Customer Data and Bespoke Configurations to create System Data and improv
No training opt-out mechanism was identified in the legal documents.
High risk across multiple dimensions. Only use with public, non-sensitive data.
deepset explicitly retains broad rights to use Customer Data and Bespoke Configurations to create System Data and improve services. Section 4.5 grants deepset a 'non-exclusive, non-transferable, sub-licensable (through multiple tiers), fully paid-up worldwide right and license to use Customer Data and Bespoke Configurations' for service provision and System Data creation. This enables training on user content.
“deepset and its Affiliates is hereby granted a non-exclusive, non-transferable, sub-licensable (through multiple tiers), fully paid-up worldwide right and license to use Customer Data and Bespoke Configurations: (a) to provide the Services and Professional Services; and (b) to create System Data.”
The license granted by customers is extremely broad. Section 4.5 grants deepset worldwide, perpetual rights to use Customer Data and Bespoke Configurations with sub-licensing through multiple tiers. The license is 'fully paid-up' and non-exclusive but allows deepset extensive exploitation rights beyond just service provision.
“deepset and its Affiliates is hereby granted a non-exclusive, non-transferable, sub-licensable (through multiple tiers), fully paid-up worldwide right and license to use Customer Data and Bespoke Configurations”
Section 4.5 explicitly permits sub-licensing 'through multiple tiers,' which enables deepset to share Customer Data with third parties. Additionally, Section 4.6 indicates third-party model providers (e.g., OpenAI) may receive data, and customers must pay fees directly to these providers, confirming data sharing occurs.
“deepset and its Affiliates is hereby granted a non-exclusive, non-transferable, sub-licensable (through multiple tiers), fully paid-up worldwide right and license to use Customer Data and Bespoke Configurations”
The agreement does not specify data retention limits or deletion rights. Section 4.5 grants perpetual rights to use Customer Data. While Section 10.3(c) requires return of materials upon termination, there is no explicit commitment to delete Customer Data or System Data, and the perpetual license suggests indefinite retention.
“deepset and its Affiliates is hereby granted a non-exclusive, non-transferable, sub-licensable (through multiple tiers), fully paid-up worldwide right and license to use Customer Data and Bespoke Configurations”
Ownership is split: Customer owns AI Applications and Bespoke Configurations (Section 4.1), but deepset retains ownership of Background IP, Foreground IP, and System Data. While customers own their outputs, deepset retains broad rights to use Bespoke Configurations and derives System Data from customer usage, creating moderate risk.
“deepset exclusively owns all right, title and interest in and to the Services and Background IP...Customer exclusively owns all right, title and interest in and to Customer Data, Customer's Confidential Information, AI Applications and Bespoke Configurations.”
The Terms reference GDPR compliance and data protection laws (Section 2.3, Data Protection Laws definition) but do not explicitly mention SOC2, HIPAA, or ISO certifications. The Privacy Policy mentions GDPR compliance but provides no evidence of third-party certifications. Moderate risk due to lack of explicit certification claims.
“deepset will maintain a security program materially in accordance with industry standards that is designed to: (a) ensure the security and integrity of data uploaded by or on behalf of Customer”
Track and measure your AI tools with Olakai's observability platform.
Explore Olakai