HAL Discovery MCP: explore HAL metadata in natural language from your favourite LLM

HALSubscribe

Written by Yutong Fei

Are you finding it difficult to build queries with HAL’s APIs? If so, HAL Discovery MCP could be the solution for you. HAL is adapting to evolving usage and interaction practices with Large Language Models (LLMs) by experimenting with an MCP server that is designed to enable AI assistants to explore the platform’s metadata.

What research is being carried out on a given topic? Which researchers and laboratories are the most active in this field? Which ANR or European projects are associated with it? What recent publications are there from a particular laboratory or researcher? Which laboratory is Y affiliated with? HAL Discovery MCP is an MCP server that allows you to query HAL using natural language via your preferred LLM (e.g. ChatGPT, Emmy for CNRS staff, Claude). Please note that these LLMs cannot access the full text of files deposited in HAL — only the metadata describing the publications (title, abstract, author, publication year, etc.).

What is an MCP server?

The Model Context Protocol (MCP), developed by Anthropic — the company behind the Claude language model — in late 2024, is an open-source standard that connects generative AI systems to various data sources (tools, databases and models) and enables them to interact with them. An MCP server exposes resources (data, tools, prompts) in a structured way that the AI can understand, assemble and use autonomously to meet users’ needs expressed in natural language.

How to use HAL Discovery MCP

The first essential step is to connect the HAL Discovery MCP server to your favourite LLM. Each AI agent has its own connection method, and the user documentation includes connection guides for some of them. However, it is worth noting that MCP usage is generally restricted to paying subscribers of LLMs such as ChatGPT, Mistral, Claude, etc. For French institutions, CNRS staff can connect the HAL Discovery MCP to Emmy, CEA staff to Maïa, and INRAE staff to Argo.

Once you are connected, it’s very simple: just ask your questions! An initial set of tools has been developed to query metadata based on publication and author information, as well as to analyse the publishing activity of research structures indexed in HAL. To ensure you receive reliable answers, it is recommended that you phrase your questions in line with the features covered by the available tools. This is why the user documentation includes a prompt library — example queries — to help you formulate your own.

Origins of the project

The HAL Discovery MCP project was initiated at the request of the CEA. Over the past two years, the CEA has collaborated with Mistral AI to develop Maïa: an experimental, sovereign and secure generative AI platform designed for the organisation’s 22,000 employees. With this in mind, the CEA is developing its own MCP servers to provide its teams with access to resources, while also planning connections with the services of trusted partners, in compliance with its data protection requirements. Against this backdrop, the CEA requested that the CCSD develop an MCP server to provide direct access to HAL metadata from Maïa.

Meanwhile, several French institutions, including the CNRS, the University of Rennes and the URFIST of Paris, asked the CCSD to consider new HAL services adapted to evolving generative AI practices. They also emphasised the importance of data security and transparent tools. These requests align with the CCSD’s ambition to maintain HAL as the benchmark infrastructure for scientific information. In an ever-changing landscape of information access, it is crucial that HAL remains accessible wherever it is used, without compromising the control and transparency standards that define the institutional environment. Against this backdrop, the development of an MCP server for HAL is intended to facilitate access to its metadata via the AI tools employed by the scientific community, thereby complementing existing APIs while providing a transparent, sovereign solution. HAL Discovery MCP enhances the discoverability of HAL’s metadata and its compliance with the FAIR principles by making it searchable via AI tools, thereby benefiting both authors and the scientific community as a whole. This initiative forms part of the HALiance project.

Help us improve HAL Discovery MCP

HAL Discovery MCP provides users with a transparent and sovereign solution that guarantees the protection of their data. As the service is still in the experimental phase, we need your feedback. Please share your experience of using HAL Discovery MCP by answering the questions in this survey. Your responses will help us to improve the service and provide better support for using LLMs to consult HAL metadata.

We are organising a webinar to introduce HAL Discovery MCP on Monday, 16 November at 13:00. Click on this link to register and receive the webinar link.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.