v2.6.0

This release adds Agentic RAG support with plan-and-execute pipelines, streaming responses, and UI integration; changes the default vector database to Elasticsearch and the default object store to SeaweedFS; adds Red Hat OpenShift support for Helm-based deployment; and introduces new agent skills for deployment, evaluation, and performance tooling.

Highlights

This release includes the following key updates:

• Added Agentic RAG support, including the plan-and-execute pipeline, streaming responses, and RAG UI integration.
• Changed the default vector database to Elasticsearch.
  • GPU accelerated support needs enterprise access and is disabled by default.
  • Milvus remains available as an optional vector database backend.
• Changed the default object store to SeaweedFS from MinIO.
• Updated the default LLM to nvidia/nemotron-3-super-120b-a12b and enabled Nemotron reasoning by default in deployment configurations.
• Promoted nvidia/llama-nemotron-embed-vl-1b-v2 as the default embedding model. The text embedding model nvidia/llama-nemotron-embed-1b-v2 remains available as an optional configuration.
• Added VLM reranker support as an opt-in.
• Added dynamic filter expression generation for Elasticsearch.
• Published RAG performance tooling and skills to use it easily.
• Published the RAG evaluation framework and skills to use it easily.
• Updated NV-Ingest to version 26.3.0.
• Updated OCR NIM naming from nemoretriever-ocr-v1 to nemotron-ocr-v1.
• Added OpenClaw plugin for agent-driven deploy/configure/eval workflows.
• Added Red Hat OpenShift and OKD support for Helm deployments.

Fixed Known Issues

The following known issues have been resolved in this release:

• Fixed default LLM sampling parameter handling for non-NVIDIA providers.