Hi Mark, I've come across various pieces of information that might answer (or begin to answer) your questions using the Primo Research Assistant as another example. The getting started<https://knowledge.exlibrisgroup.com/Primo/Product_Documentation/020Primo_VE/...> documentation talks about using RAG architecture<https://aws.amazon.com/what-is/retrieval-augmented-generation/> to reduce risks of hallucinations and the AI privacy policy<https://clarivate.com/ai/academia/policy/> linked mentions that "Clarivate does not use your data to directly or indirectly train LLMs." So, take that how you will! They also mention using OpenAI's GPT--4o mini for the current AI model. Clarivate hasn't mentioned a specific threshold for excessive use for the PRA, but they do say "Currently, there are no restrictions on the number of searches. However, we are monitoring usage to control costs for the Large Language Model and may need to set a limit on the number of searches an institution can perform within a certain timeframe in the future." Best, Alison ////////////////////////////////////////////// Alison L. Davis Discovery and Integrated Library Systems Administrator Purdue University Libraries and School of Information Studies 504 Mitch Daniels Blvd | West Lafayette, IN 47907 lampley@purdue.edu<mailto:lampley@purdue.edu> | Schedule a Teams meeting with me!<https://calendar.lib.purdue.edu/appointments/lampley/teams> From: Erin Nettifee (she/her/hers) <erin.nettifee@duke.edu> Sent: Thursday, April 10, 2025 10:30 AM To: Kluzek, Mark <m.kluzek@ucl.ac.uk>; ai-sig@exlibrisusers.org Subject: [Ai-sig] Re: Ex Libris & AI - technical documentation You don't often get email from erin.nettifee@duke.edu<mailto:erin.nettifee@duke.edu>. Learn why this is important<https://aka.ms/LearnAboutSenderIdentification> ---- External Email: Use caution with attachments, links, or sharing data ---- Hi Mark, Using the Metadata Assistant as an example, they are not getting into technical details, but they are giving some more in-depth responses in the FAQ: https://knowledge.exlibrisgroup.com/Alma/Product_Documentation/010Alma_Onlin... "Q: What LLM are you working with? A: Models are subject to change as models and capabilities evolve. As of the February 2025 release, the AI Metadata Assistant was working with the gpt_4o_2024_08_06 LLM model." Best, Erin Nettifee Duke University Libraries ________________________________ From: Kluzek, Mark Sent: Thursday, April 10, 2025 10:18 AM To: ai-sig@exlibrisusers.org<mailto:ai-sig@exlibrisusers.org> Subject: [Ai-sig] Ex Libris & AI - technical documentation Hi all, I've been wondering whether there are any technical documents that outline, in some depth, how AI works with Ex Libris products. I'm thinking along the lines of the documentation you see in the Developer Network. Specifically, I was hoping it might address questions such as: * What specific AI models are used for the different AI-related functionalities in Alma? Architectural diagrams outlining the flow of data would be useful. * Is there a usage threshold in the event of excessive use? * What risks are there of AI "hallucinations"? How do the models ensure the greatest possible accuracy of information? * In terms of privacy, is the data we input retained or used in any way? I've engaged a little with Ex Libris, as well as with other members of the Ex Libris user community. I don't believe anything like this currently exists, but I do think there would be value in greater transparency around how AI is being used. I'd be interested to hear if anyone else has similar thoughts or insights on the matter. Best regards, Mark