Generative Artificial Intelligence Laboratory for Agent-based Search Orchestration – LIAGORA

CentraleSupélec's MICS laboratory and ILLUIN TECHNOLOGY are joining forces to create a LabCom. Its aim is to enable the deployment and use of Generative AI within companies, while guaranteeing the use of trusted and frugal systems. Capable of creating original content, formulating knowledge-based answers to complex questions, and acting as a versatile assistant for a wide range of professions, generative AI offers considerable potential for optimizing business processes. In particular, ILLUIN is developing a suite of products capable of processing document data (search engine for complex document corpora, document parser) or conversational data (conversational agents, conversation parser, corpus analyzer), as well as a multimodal orchestration platform that enables the design of customized AI use cases.
However, the limitations of current assistants based on generative AI are still too numerous. They are too prone to hallucinations, resulting in inaccurate or incoherent responses. They use a RAG (Retrieval Augmented Generation) system to answer many queries, yet the RAG approach itself has many limitations, particularly in information retrieval for complex queries within vast specialized documentary corpora. They generally lack long-term context, which can lead to omissions or contradictions in extended conversations. They too rarely make use of the company's business rules, which can be formulated in a variety of structures (documents, ontologies, databases, etc.). Lastly, they require infrastructures that are too large for their widespread use to be acceptable to most companies.
The three major challenges of this LabCom will therefore be to develop assistants based on Generative AI:
- Efficient, by improving their relevance and diversifying possible use cases,
- Trustworthy, by reducing hallucinations through optimization of the foundation models used (LLM, VLM, etc.) or information retrieval paradigms (RAG, Agents, use of Tools, etc.),
- Frugal, to reduce the carbon, energy and financial impact of the systems developed (through hybridization, use of small foundation models, model compression, etc.) to enable mass use and adoption with reasoned consumption.