BNL Home
|
Directory
BNL Tech Notes
Toggle navigation
Ernest: A Retrieval-Grounded AI Assistant for EIC Accelerator Physics
Tech Note Information
Name
Electron-Ion Collider Project (EC)
BNL Number
BNL-230305-2026-TECH
BNL Series Number
230305
Description/Abstract
The EIC project’s technical knowledge, including lattice files, technical notes, controlled documents, presentations and simulation code, is large and spread across many repositories and formats. General-purpose language models answer questions about it fluently but without grounding, so a plausible but wrong parameter value can look exactly like a correct one. This note describes Ernest, an assistant built to close off that failure mode. Ernest pairs an agentic language model with (i) a local, retrieval-only knowledge server exposed over the Model Context Protocol (MCP), (ii) an audited, provenance-carrying snapshot of canonical per-ring parameters for the modelled accelerator chain, (iii) skills, which package validated workflows that Ernest follows\\\\r\\\\nand can learn, kept separate from the searchable index, (iv) memory tiers organized by lifetime, from a crash-safe per-project terminal history to a shared, git-versioned ledger of learned facts, and (v) direct access to accelerator modelling codes (MAD-X, Bmad/Tao, SciBmad) so it can check numerical claims computationally. A single launcher, ernest claude|codex|opencode, starts it from any working directory with any of three agent harnesses. The central design rule is that every answer separates documented facts, calculated results, hypotheses and model inferences. We describe the architecture, show representative queries and a live lattice-building demonstration, and review four campaigns of RCS lattice work carried out with the assistant between July and September 2026, which ended in the adoption of RCSV6S0 as the new 18 GeV baseline. We close with limitations, open questions and future directions.
Publication Date
9/30/2026
Submittal Date
10/1/2026
Author Information
Primary Author
V. Ranjbar
DOE Subject Category
PARTICLE ACCELERATORS
PDF File
Back to Results