Work

AskALCF: A RAG Assistant for HPC User Support

Year: 2026
Type: article
Abstract: Traditionally, HPC user support has relied on two main channels: written user guides and direct help from facility staff. That model is under pressure as supercomputing systems become more complex, heterogeneous, and dynamic. At ALCF, users must navigate Aurora, Polaris, inference services, scheduler policies, programming environments, and operational status information spread across multiple documentation pages and service sites. At the same time, the users of these systems now include not only scientists and Software Engineers, but also AI agents that need to retrieve guidance, reason about it, and act on behalf of users. This raises a practical challenge for the AI era: how can facilities use AI to make user support faster, more accurate, and more scalable while keeping answers grounded in trusted documentation and operational context? AskALCF is a production retrieval-augmented generation (RAG) assistant built to address this challenge. It indexes public user docs from ALCF, NERSC, OLCF, and other DOE facilities into a vector store that preserves images and code blocks. Two front ends share the same retrieval pipeline: a Streamlit web app for end users, and a Model Context Protocol (MCP) server for agent clients such as Claude Code, Codex, and IDE assistants. A second MCP server exposes read-only HPC operations (system status, allocations, IRI facility API), so an agent can combine docs and live cluster state in one tool call. The talk covers three things. First, the architecture of AskALCF and how we run the system in production: lab authentication, a cross-facility document index refreshed daily as the upstream guides change, and the day-to-day of keeping the service stable. Second, what we learned from a year of real usage — who asks, what they ask, where the chatbot still fails, and what the upgrades actually moved. Third, the roadmap to an autonomous agentic user support system including a planned ticket-solving agent and an offline trace-mining loop that would draft documentation pull requests with verify-by-execution evidence. The earlier ChatBot underlying this work was published at the SC25 HUST workshop [1]. This talk focuses on our ongoing efforts since then: the dual-MCP interface, production deployment, the measured impact on real users, and the agentic roadmap. We aim to present these contributions and invite comparison with related work at other HPC centers.
Language: English
Open Access status: green
OpenAlex ID: W7213512566
FWCI:
Cites:
Cited by:
Related to: 10