Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Proposal of an AI-Based Support Assistant for the ALICE-FIT Detector Setup at CERN

Ignacy Mermer, Jakub Muszyński, Jakub Możaryn, Krystian Rosłon

TL;DR

The paper tackles the challenge of rapid, reliable decision-making in the ALICE-FIT DCS under time pressure and information fragmentation by proposing an AI-assisted system that combines a retrieval-augmented LLM with a curated, versioned ALICE documentation corpus. It deploys an agentic RAG framework using Self-RAG and ReAct-style reasoning, augmented with constitutional AI safeguards and RBAC to ensure safety, auditability, and minimal hallucinations. The architecture integrates with the existing ALICE control stack (ALF/IPbus, FRED, WinCC OA, DIM) in an air-gapped environment, enabling read-only diagnostics initially and controlled, gated actions via Quick Commands. Evaluation focuses on a blend of technical KPIs (accuracy, latency, MTTD, interventions) and human-centric metrics (perceived workload, trust), aiming to demonstrate improved operational efficiency and reliability with potential to scale to multiple detectors for Run 5.

Abstract

We propose an AI-based assistant designed to support the ALICE Fast Interaction Trigger (FIT) detector operators at CERN. The assistant helps diagnose and resolve operational issues in the Detector Control System (DCS), where decisions must often be made quickly and with incomplete information. By combining Large Language Models (LLMs) with a controlled Retrieval-Augmented Generation (RAG) pipeline, the system can generate context-aware suggestions based on verified ALICE-FIT documentation and problems that have appeared in the past.

Proposal of an AI-Based Support Assistant for the ALICE-FIT Detector Setup at CERN

TL;DR

The paper tackles the challenge of rapid, reliable decision-making in the ALICE-FIT DCS under time pressure and information fragmentation by proposing an AI-assisted system that combines a retrieval-augmented LLM with a curated, versioned ALICE documentation corpus. It deploys an agentic RAG framework using Self-RAG and ReAct-style reasoning, augmented with constitutional AI safeguards and RBAC to ensure safety, auditability, and minimal hallucinations. The architecture integrates with the existing ALICE control stack (ALF/IPbus, FRED, WinCC OA, DIM) in an air-gapped environment, enabling read-only diagnostics initially and controlled, gated actions via Quick Commands. Evaluation focuses on a blend of technical KPIs (accuracy, latency, MTTD, interventions) and human-centric metrics (perceived workload, trust), aiming to demonstrate improved operational efficiency and reliability with potential to scale to multiple detectors for Run 5.

Abstract

We propose an AI-based assistant designed to support the ALICE Fast Interaction Trigger (FIT) detector operators at CERN. The assistant helps diagnose and resolve operational issues in the Detector Control System (DCS), where decisions must often be made quickly and with incomplete information. By combining Large Language Models (LLMs) with a controlled Retrieval-Augmented Generation (RAG) pipeline, the system can generate context-aware suggestions based on verified ALICE-FIT documentation and problems that have appeared in the past.

Paper Structure

This paper contains 8 sections, 2 figures.

Table of Contents

  1. Introduction and Objectives
  2. Literature Review and Methods
  3. System Design and Configuration
  4. Knowledge-Base Governance and Corpus Curation
  5. From Read-Only Advice to Safe Action
  6. Implementation
  7. Testing and Evaluation
  8. Summary and Further System Development

Figures (2)

  • Figure 1: Control and monitoring architecture. AI-Based Support Assistant and WinCC exchange data bidirectionally via DIM and connect to FRED. FRED bridges to the ALF/IPbus--ALF layer over DIM, which interfaces with the front-end electronics through GBT/IPbus.
  • Figure 2: Overview of the evaluation framework. The methodology integrates quantitative technical KPIs (e.g., accuracy, MTTD) with qualitative human-centred metrics (e.g., perceived workload, trust) to provide a holistic assessment of the AI assistant's impact on detector operations.