SAGE builds open-source systems for agent, retrieval-augmented generation, workflow, and service orchestration.
Our work spans agent coordination, tool use, RAG pipelines, evaluation, safety, application-level state semantics, visual workflow authoring, tutorials, and reproducible application benchmarks.
Website: sage.org.ai
- IntelliStream incubates early research ideas across the ecosystem.
- SAGE owns application, agent, RAG, workflow, evaluation, and service-orchestration systems.
- DataSys provides framework-neutral stream, graph, vector, index, update, query, and benchmark infrastructure.
- vLLM-HUST develops model execution, KV/cache scheduling, compilation, kernels, and hardware acceleration.
- SAGE - application, agent, RAG, workflow, and service orchestration platform.
- SAGE Docs - documentation and the source for sage.org.ai.
- SAGE Agentic
- SAGE RAG
- SAGE Eval
- SAGE Studio
- SAGE Tutorials
- SAGE Agent for VS Code
- SAGE Agentic Tool Use
- SAGE Examples
- SAGE Finetune
- SAGE GitHub Manager
- SAGE Intent Library
- SAGE Mate
- SAGE Refiner
- SAGE PyPI
- SAGE Wiki
- SAGE Benchmark
- Agentic Tool-use Benchmark
- Agentic Tool-use SIAS
- RAG Benchmark
- Refiner Benchmark
- Wiki-link Retrieval
- Neuromem Benchmark
- Complete publication timeline - published work, public evidence, corresponding repositories, and explicitly labeled author-reported acceptances.
- SAGE - ICML 2026 dataflow-native framework for modular, controllable, and transparent LLM-augmented reasoning.
- Neuromem - ICML 2026 benchmark and lifecycle decomposition for external memory in LLM systems.
- FlowRAG - WWW 2026 continual retriever adaptation for evolving corpora.
- Demonstrating SAGE - ICPP 2026 demo acceptance reported by the authors; the public program or proceedings record is pending.
Projects graduate from IntelliStream when they have a clear public abstraction, named maintainers, reproducible evaluation, documentation, tests, licensing, and a sustainable release path.
SAGE is not a data-engine or model-runtime umbrella. Framework-neutral data systems belong in DataSys, while model-runtime and hardware-execution projects belong in vLLM-HUST.
See our contribution guide, security policy, and support guide.