Case Study

Albert

The first LLM-driven assistant for French public servants

THE DIRECTION INTERMINISTÉRIELLE DU NUMÉRIQUE (DINUM), IN COLLABORATION WITH THE FOUNDING TEAM OF PLEIAS, LAUNCHED THE DEVELOPMENT OF ALBERT, AN ECOSYSTEM OF GENERATIVE AI TOOLS DESIGNED FOR PUBLIC SERVANTS. THE INITIATIVE AIMS TO ENHANCE THE EFFICIENCY AND CUSTOMIZATION OF PUBLIC SERVICES THROUGH ADVANCED LLM CAPABILITIES IN DOCUMENT PROCESSING, INFORMATION RETRIEVAL AND TEXT GENERATION.

THE DIRECTION INTERMINISTÉRIELLE DU NUMÉRIQUE (DINUM), IN COLLABORATION WITH THE FOUNDING TEAM OF PLEIAS, LAUNCHED THE DEVELOPMENT OF ALBERT, AN ECOSYSTEM OF GENERATIVE AI TOOLS DESIGNED FOR PUBLIC SERVANTS.

THE INITIATIVE AIMS TO ENHANCE THE EFFICIENCY AND CUSTOMIZATION OF PUBLIC SERVICES THROUGH ADVANCED LLM CAPABILITIES IN DOCUMENT PROCESSING, INFORMATION RETRIEVAL AND TEXT GENERATION.

This project focuses on introducing a knowledge assistant to help HR teams to navigate the intricate regulations and documentation that govern the varied personnel types within the education sector, and bring appropriate answers to their queries.

Objectives

Personalized Responses:

text

Albert is engineered to provide tailored assistance to public servants, enabling them to address citizen queries with high precision.

Empowerment through Technology:

text

By automating routine tasks, Albert aims to allow public servants to focus on higher-value activities, including personalized citizen support.

Transparency and Accessibility:

text

The tool prioritizes clear sourcing and easy access, ensuring that it can be seamlessly integrated across various administrative departments.

Challenges and Solutions

Customization Needs:

text

Current open weight LLMs, while proficient in basic interactions, lack the specificity and reasoning capacities required for public service applications. Through custom fine-tuning strategies involving administrative data as well as synthetic reasoning data sets, Albert was enhanced to better answer these specialized needs.

User Training:

text

Recognizing the importance of more conscious and deliberate interactions with LLMs, training programs for public servants on understanding generative AI functioning as well as on crafting efficient prompts for Albert were initiated.

Technology and Methodology

text

Albert is currently built on open weights LLMs, such as Mistral 7b and LLaMA-7b, chosen for their reasoning and conversational capacities doubled with energetically efficient inference due to models size.

text

The development process involved fine-tuning these models on specific datasets relevant to public service regulations and queries, developing Retrieval-Augmented Generation pipelines as well as interpretability system for citing retrieved references.

Deployment

text

Albert has been rolled out within the France services network, initially targeting volunteer advisors. This step marks the beginning of a broader implementation strategy that aims to make Albert available to all state administrations in 2024.

text

Though only in its pilot phase, Albert demonstrates potential in streamlining administrative processes, offering a scalable solution to enhance public service delivery.

CONTACT@PLEIAS.FR