v2.26.0-rc1

Deepset has launched Haystack v2.26.0-rc1, a significant pre-release update to its popular open-source framework for building production-ready LLM applications. The headline feature is the new LLMRanker component, which represents a major advancement in retrieval-augmented generation (RAG) pipelines. Unlike traditional keyword-based ranking, LLMRanker uses a ChatGenerator and PromptBuilder to rerank documents based on JSON-formatted LLM output, allowing for much more nuanced and context-aware document prioritization. This enables developers to build systems that understand semantic relevance rather than just keyword matching.

The release also brings official support for OpenAI's newest models, including the recently announced GPT-5-mini and GPT-4o, through both OpenAIChatGenerator and AzureOpenAIResponsesChatGenerator components. This gives developers immediate access to cutting-edge model capabilities without waiting for framework updates. Additionally, the Agent component now supports Jinja2 templating for system prompts, allowing dynamic variable injection at runtime—perfect for incorporating real-time data like current time or location into agent instructions.

Other notable improvements include enhanced performance for HuggingFaceAPIDocumentEmbedder through concurrent embedding requests, Python 3.14 compatibility, and better tool customization in SearchableToolset. The update also fixes several technical limitations, such as allowing components with union-type inputs (like list[str] | list[ChatMessage]) to accept multiple connections, which was previously restricted. These changes collectively make Haystack more flexible, performant, and ready for next-generation AI applications.