Albert, the innovative generative AI, is reinventing public administration. *This revolutionary device* deploys remarkable applications, *deeply transforming the daily life of agents.* The State relies on this technology to modernize its services while ensuring *a transparent interaction with citizens.* By combining the power of language models and open source, Albert shapes a bold digital future while addressing sovereignty challenges.
The large language model (LLM) named Albert, introduced in April 2024, represents a significant advancement for the State as well as for its institutions. Initially designed to assist agents in their administrative tasks, Albert is now being rolled out on a large scale within ministries, particularly in the Justice and Culture ministries, as well as within the Academy of Lyon.
Based on the open source model Llama 3.1 from Meta and other LLMs from Mistral, Albert embodies a sovereign AI solution. This technology, completely open source, aims to ensure the security and reliability of sensitive data handled within public administration. Ulrich Tan, head of the Datalab, emphasizes that this initiative has been established to address each use case in generative AI while ensuring that the tool is exclusively intended for public agents.
Features of Albert
Albert encompasses a wide range of generative AI capabilities. Its design allows for varied functionalities, ranging from automatic writing to document synthesis, not to mention augmented retrieval generation (RAG). This latter process allows for extracting answers from a specific corpus of files, thus improving the quality of information provided to agents.
A notable feature of Albert lies in its use as an assistant to support agents, maintaining responsibility for the relationship with users. Agents benefit from powerful tools while being involved in managing interactions with citizens.
Implementation by the State
The interministerial digital directorate (Dinum) has developed Albert in a customized approach, thus allowing ministries to access its capabilities through an API. This initiative has been named LLM as a Service, thus offering the 2.5 million state agents the opportunity to integrate these technologies into their daily activities. The Dinum operates under a structured and open deployment logic, having already made Albert’s code accessible on platforms such as GitHub and Hugging Face.
Experiments and initial results
At the heart of a state start-up incubated by the Dinum accelerator, Albert has been integrated within France Services. This experiment gave rise to Albert France Services, designed to support state advisers in responding to citizens’ administrative inquiries. Preliminary results are promising: 71% of users report ease of use, and 58% wish to recommend the tool to their colleagues.
Another use case for Albert is its connection with compar:IA, a conversational AI comparator developed by the Ministry of Culture. Albert is also integrated within the state’s office suite, where it assists with rephrasing, synthesis, and text generation, thus facilitating the work of administrative agents.
Investments and evolution prospects
The State has invested over one million euros in the development of Albert during the year 2024. The projects related to Albert, due to their open source nature, aim to promote open innovation within administrations. The gradual implementation of Albert’s capabilities takes place in several stages, with a clear objective: to design production applications in short deadlines, thus avoiding the pitfalls of previous complex IT projects.
The Dinum is considering research and development projects around a knowledge graph that will ensure more precise answers, particularly on legal questions. This advancement could reduce the errors that some current models may generate, thus improving the quality of the information provided.
Multi-agent RAG and new innovations
A research project based on Albert is exploring the possibility of a multi-agent RAG. This device will allow multiple agents to assess and process the questions posed, classifying requests according to their complexity. A synthesis work will then be performed to provide the most relevant answers.
Societal implications
Albert is not limited to just improving administrative efficiency. Its implementation also raises questions about the ethics and governance of AI technologies. The multiplication of projects must occur within a framework of transparency and accountability to avoid risks associated with inequalities and the digital divide.
Albert’s ambitions are set within a context of increasing regulation and security, both at the national and European levels. Initiatives like the future European regulation of generative AI will rely on tools like compar:IA, thus illustrating the potential of this technology to enhance administrative capabilities while ensuring user protection and fairness.
Frequently asked questions about Albert, the generative artificial intelligence of the State
What is Albert and what is its main function within the administration?
Albert is a large language model (LLM) developed by the interministerial digital directorate to assist public agents in their missions. It is designed to facilitate tasks such as automatic text writing and document synthesis.
How is Albert integrated and used by the various ministries?
Albert is made available to ministries via an API, allowing them to develop customized applications to meet their specific needs in generative artificial intelligence.
What are the implications of Albert for state agents and their relationship with users?
Albert is designed as an aid for agents, who remain responsible for interaction with users. This allows for the automation of certain tasks while maintaining human responsibility in the process.
What concrete use cases have been developed with Albert so far?
Use cases include supporting advisers from France Services in responding to citizens’ administrative inquiries, as well as its integration into office tools for text generation and synthesis.
How does Albert manage sensitive data and information security?
Albert is installed on instances under full control of the administration, ensuring the protection of sensitive data processed. Its open-source model also allows for addressing various use cases while respecting information security.
What feedback has been received on the use of Albert among advisers from France Services?
Initial feedback indicates that 71% of users find Albert easy to use, and 58% declare themselves ready to recommend the tool to their colleagues, highlighting its effectiveness in the context of administrative missions.
What is the State’s commitment regarding financing and developing Albert?
For the year 2024, the State has invested over one million euros in the development of Albert, with an intention to continue opening and sharing the resulting projects under an open-source approach.
Is Albert intended only for the French administration or can it be used internationally?
Although Albert is primarily developed for the French administration, the model can inspire other public administrations around the world, and its open-source model could be adapted to similar needs elsewhere.
What future innovations are planned for Albert and associated language models?
Future projects include integrating a knowledge graph to improve answer accuracy, as well as developing a multi-agent architecture to manage complex questions more effectively.
