Communication Dans Un Congrès Année : 2025

Perspicacité-AI

Résumé

Recently, emerging tools such as LLM have gained attention in scientific research and education. However, their limitations, notably outdated knowledge and untrackable provenance, hinder their ability to keep pace with rapid advancements in scientific fields, like metabolomics. Some of these challenges can be addressed by coupling LLM with Retrieval-Augmented Generation (RAG) techniques, which combine relevant data retrieval with generative AI. Here, we developed Perspicacité-AI, a free and open source agentic literature assistant with advanced RAG capabilities designed for scientific research and education. Perspicacité-AI required the development of the Bibt2KB python package, which seamlessly transforms open-access bibliographic references into a structured and searchable Facebook AI Similarity Search (FAISS) knowledge base (KB). The generated KB is used by the pipeline’s framework to deploy an AI-assistant enhanced by RAG, ensuring that each question asked is grounded in relevant literature and documents. The framework features different advanced modes of document search and question-answering, including Perspicacité-Profound, an iterative, structured reasoning workflow suitable for answering complex scientific questions with the trackable facts and sources. Furthermore, the pipeline supports major open source and proprietary LLM providers, broadening the access for different users. Perspicacité-AI also incorporates a novel reranking function for document selection, Sigmoid Weighted RRF (SW-RRF). The benchmarking of SW-RRF showed means of up to 4.5% improvement over the traditional Reciprocal Reranking Function (RRF) for all tested metrics: Recall, Precision, F1, Mean Reciprocal Rank (MRR), Average Precision (AP) and Normalized Discounted Cumulative Gain (NDCG) at 1, 3, 5 and 10 documents retrieved at no computational expenses. Initial applications of Perspicacité-AI in computational metabolomics [metaboguide.holobiomicslab.eu] demonstrated effective chatbot responses with accurate citations, and while still ongoing, initial evaluations of the pipeline show promising results, not only mitigating some of the limitations of standalone LLMs but empowering scientific literature with generative AI.

Fichier principal
Vignette du fichier
Abs_MetabolomicsCoteDAzur.pdf (102.94 Ko) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-05233151 , version 1 (01-09-2025)

Licence

Identifiants

  • HAL Id : hal-05233151 , version 1

Citer

Lucas Pradi, Tao Jiang, Matthieu Féraud, Madina Bekbergenova, Yousouf Taghzouti, et al.. Perspicacité-AI. Metabolomics Symposium, Jul 2025, Nice, France. ⟨hal-05233151⟩
157 Consultations
139 Téléchargements

Partager

  • More