Researchers presented the AquiLLM platform for working with scientific data

A group of scientists presented AquiLLM — a specialized platform with open weights designed for deploying RAG (Retrieval-Augmented Generation) systems in an academic environment. The system is focused on working with closed knowledge bases and scientific archives, providing local control over data infrastructure.

The development is based on the concept of "faithfulness" of answers, defined as the degree of compliance of generation with the provided scientific context without invented facts. The authors tested the platform on astronomy tasks, evaluating the model's ability to extract information from profile documents and form analytical conclusions.

The research results showed that AquiLLM demonstrates high reliability when performing direct search queries based on specific fragments from the knowledge base. However, when synthesizing data or resolving ambiguous formulations, the model's accuracy decreases, requiring additional verification by experts.

The authors emphasize that standard benchmarks do not always reflect the specifics of scientific work, so a domain-specific approach is necessary to evaluate such systems. The platform is distributed under the CC BY-SA 4.0 license, which allows research groups to adapt the solution to their own needs.

The development is aimed at solving privacy problems when working with sensitive data that cannot be transmitted to cloud APIs. The use of local models allows institutions to maintain full sovereignty over their archives and research results.