MyGPT: Open-Source AI Platform Brings Private, Local Knowledge Search to Sensitive Data

Large language models have reshaped how researchers, clinicians and policymakers search for information, but their adoption in sensitive settings has been held back by two persistent problems: the risk that confidential data leaks to external servers, and the tendency of models to invent plausible-sounding answers when asked about material outside their training data. A team led by Jaimin Patel and M. Madan Babu now presents MyGPT, a fully open-source platform described in Nature Communications that is designed to tackle both issues at once.

MyGPT is built around retrieval-augmented generation, a technique in which the model does not answer from memory alone. Instead, when a user poses a question, the system first retrieves relevant passages from a user-defined library of documents and then generates an answer grounded in those retrieved texts. The library can contain biomedical literature, confidential clinical documents or multilingual global health policies, and the platform runs locally or within secure institutional environments rather than on external cloud services.

The privacy architecture is central to the design. Because the underlying models are open-source and the entire pipeline operates inside the user’s own infrastructure, sensitive data never leaves the secure environment. The authors position this as a direct response to the privacy concerns and data-leakage risks that have limited the use of large language models in sensitive domains such as science, clinical work and policy.

To address reliability, the authors introduce confidence metrics that accompany every generated answer. These include a relevance score, which indicates how closely the retrieved material matches the query, and a hallucination index, which is intended to flag answers that lack adequate support in the library. The interface also provides explicit citations, showing which passages from the library underpin each response, so users can verify the contextual support for themselves within a single view.

These metrics are a design feature of the platform rather than a demonstrated guarantee of accuracy; the paper describes them as tools for assessing contextual support, and users are expected to interpret them alongside the cited sources. The system is therefore positioned as an aid to careful knowledge-seeking, not as an autonomous authority whose outputs can be accepted uncritically.

Cost is another stated motivation. By combining open-source models with local deployment, the authors describe MyGPT as a cost-effective, modular alternative to commercial LLM services, one in which institutions can select their own components and adapt the system to their needs while keeping data and computation under their own control.

The authors emphasize democratization as the broader goal. Because the platform can run without internet access, it is intended to be usable at remote facilities that lack reliable connectivity, extending AI-assisted knowledge-seeking to settings that depend on locally hosted systems. The modular design also means the system can be adapted across multiple domains beyond biomedicine.

As a platform description, the paper outlines the architecture, the confidence metrics and the deployment model rather than reporting clinical or performance outcomes in specific patient or policy settings. Even so, the work addresses a recognized gap: the same privacy constraints and hallucination risks that have limited LLM use in science, medicine and policy are the constraints the platform is explicitly engineered around, offering institutions a route to AI-assisted knowledge-seeking that keeps control, data and cost within their own hands.

Subject of Research: Development of an open-source, privacy-preserving retrieval-augmented generation platform for reliable knowledge-seeking with large language models

Article Title: Democratizing reliable knowledge-seeking with MyGPT: a privacy-first, open-source retrieval-augmented generation platform

Article References: Patel, J., Downing, J., Ferguson, H., You, T., Malinverni, D., D. A, S. M., Chen, I., Sluter, M., Moorefield, B., Paréj, K., Ragavan, M., Morris, C., Keerthi, D., Becerra Armada, D., Mészáros, B., Trivedi, V. D., Alam, S., Woodard, A., Alford, D., … Babu, M. M. (2026). Democratizing reliable knowledge-seeking with MyGPT: a privacy-first, open-source retrieval-augmented generation platform. Nature Communications. https://doi.org/10.1038/s41467-026-78509-6

Image Credits: AI Generated

DOI: 10.1038/s41467-026-78509-6

Keywords: large language models, retrieval-augmented generation, open-source software, data privacy, hallucination, biomedical literature, clinical documents, global health policy, knowledge discovery, MyGPT, secure computing, Nature Communications