Speaker
Description
This talk presents the development of an agentic chatbot for the LHCb experiment, a project realized in cooperation with ItGPT, the AI Chatbot collaboration at CERN. The assistant is intended to support learning, operations, software development, and data analysis tasks.
The LHCb knowledge base is structured in three access tiers: public, CERN-shared, and internal. The internal knowledge includes both raw data and the operational knowledge needed to use software tools, interpret specialized data formats and machine logs, and understand experiment-specific jargon. We will describe our progress on data preprocessing, chunking, and embedding strategies, covering non-trivial use cases such as plot generation, online operations, and detector monitoring.
The chatbot uses a Retrieval-Augmented Generation (RAG) pipeline, developed on top of AccGPT, to ground the LLM's reasoning in the LHCb knowledge base. We are now extending this system with agentic functions managed by an LLM-based intelligent router, which we call the "LHCb Brain." This router directs queries to a selection of predefined routes and Model Context Protocol (MCP) servers, which provide pathways to internal services. This architecture enables the chatbot to perform reasoning via repeated retrieval and to execute tasks, not just generate text.