Orange, in search of innovation, initiates a revolutionary project of multi-agent RAG systems, aiming to redefine artificial intelligence. This unprecedented approach will combine multiple retrieval agents, thus optimizing the analysis of various documents such as text, images, and graphics. The project promises to surpass the current limits of traditional solutions, where a single agent is often insufficient. The stakes are clearly outlined: to improve vectorization and the understanding of complex data while creating a more efficient and integrated AI ecosystem.
An innovative project at Orange Business
The specialized subsidiary of Orange dedicated to digital services for businesses is making notable efforts in developing a project of RAG (Generative Augmented Retrieval) that integrates a multi-agent approach. This ambitious project aims to transcend the current limitations of artificial intelligence systems in this area, which have so far specialized in handling a single agent.
Project Objectives
Orange Business aims to optimize the management of complex documents, combining different types of media such as text, images, tables, and graphics. The goal is to develop a solution capable of generating more precise and relevant responses from an enriched document base. The establishment of this multi-agent RAG system is designed to lead quickly to the creation of a commercial offer in the coming months.
The challenges of existing systems
Traditional RAG solutions suffer from notable limitations, exploiting only the textual content of documents. The vectorization of reference documents often presents inconsistencies, making it difficult to grasp the overall meaning. Didier Gaultier, head of AI at Orange Business Digital Services, emphasizes that users, accustomed to keyword searches, do not fully leverage the capabilities of generative artificial intelligence.
Architecture of the multi-agent RAG
The concept of the multi-agent RAG is based on the collaboration of several agents, each specialized in the vectorization of a specific type of content. These agents process various formats such as text, images, and tables, creating a unique document base, referred to as “meta-vectorized”. The development of an orchestration engine will manage the interaction between these diverse agents, ensuring improved and relevant search results.
The networking of documentary contents, through a graph system, will result from matching analyses between files. This method will allow for understanding existing connections, thus offering information retrieval that goes beyond simple vectors.
Potential use cases
This system could revolutionize how businesses manage their data. For example, it would be possible to detect that a skill of an employee, identified by the vectorization of their professional content, is not mentioned in their CV, thus facilitating the matching of skills and job titles.
At the heart of this technology, agents will collaborate to formulate intelligible phrases from keywords, thus optimizing the relevance of user queries.
Collaboration and future development
Orange Business has formed a consortium of players, including Lighton, an expert in generative AI. The desire to achieve a marketable product in the near future is evident. In collaboration with industrial partners, the IT services company is seeking financial support to accelerate the project.
This initiative promises to accelerate innovations in the field of multi-agent systems, positioning Orange as a reference player in the sector of AI solutions for businesses.
Frequently Asked Questions about Orange’s multi-agent RAG project
What is the multi-agent generative augmented retrieval (RAG) system proposed by Orange?
The multi-agent RAG is an innovative system that uses multiple agents to retrieve and process information from multimedia documents, such as texts, images, or tables, in order to provide more precise and relevant answers.
What are the advantages of the multi-agent approach compared to traditional RAG systems?
Unlike traditional systems that rely on a single agent, the multi-agent approach allows for the vectorization of different formats, which improves information retrieval by taking into account the interrelations between various elements of content.
What does vectorization involve in Orange’s RAG project?
Vectorization involves transforming the information from a document into numerical representations, enabling agents to search for and identify the most relevant vectors related to a question asked by the user.
How does Orange plan to ensure the quality and accuracy of the answers provided by the multi-agent RAG?
Orange aims to create a meta-vectorized document base that analyzes matches between files and establishes relationship graphs to ensure more refined and relevant information retrieval.
What is the timeline for the development and commercialization of the multi-agent RAG system?
Orange Business plans to launch a packaged offer of this system in the coming months, after forming a consortium with specialized partners, such as Lighton, to accelerate the project.
How will users interact with the multi-agent RAG system once it is launched?
Users will be able to ask questions in the form of complete sentences, and the multi-agent RAG system will transform these requests into intelligible prompts to perform precise searches in the database.
What steps are necessary before the multi-agent RAG project becomes operational?
Before becoming operational, the project requires the development of technical infrastructure, the implementation of retrieval agents, as well as thorough testing to validate the accuracy of the results provided.
What types of agents will be used in Orange’s multi-agent system?
The system will integrate several types of agents specialized in the vectorization of different information formats, such as text, images, tables, and potentially even audio content.
Why is it essential to have an orchestration engine in this multi-agent system?
The orchestration engine is crucial as it will coordinate the interaction between the different agents, ensuring that the responses are based on an integrated and coherent understanding of the retrieved data.
