Demonstrating SAGE: A Dataflow-Native Framework for Modular, Controllable, and Transparent LLM-Augmented Reasoning
TL;DR
This demo presents SAGE as a pipeline-native runtime system and demonstrates it through an OPC-facing control plane and includes a compact distributed comparison against a LangChain RPC baseline over a shared 15-workload RAG suite, where SAGE nearly doubles full-RAG throughput under matched 8-node settings.
Abstract
Modern LLM applications combine retrieval, generation, memory, tool invocation, and stateful services, making their runtime behavior difficult to inspect and control when assembled from loosely connected services. This demo presents SAGE as a pipeline-native runtime system and demonstrates it through an OPC-facing control plane. SAGE represents applications as declarative operator pipelines and compiles them into runtime-managed stage bindings on a Flownet-based substrate. The OPC UI exposes ExecutionGraph-derived specifications, metrics, lifecycle traces, and logs rather than a dedicated operator-level graph browser. The control plane discovers more than one hundred runnable entries, surfaces launch contracts, and lets an operator validate, launch, and invoke representative requests through a uniform UI. Result bundles, instance views, compatibility checks, metrics artifacts, logs, and experiment reports connect each UI-level entry to underlying SAGE execution. We use an OPC support-and-automation scenario as the live context and include a compact distributed comparison against a LangChain RPC baseline over a shared 15-workload RAG suite, where SAGE nearly doubles full-RAG throughput under matched 8-node settings.