Knowledge graphs as a structured memory for collaborative agents

INRIA; Valbonne; Sur place; EUR 20 000 - 40 000; Plein temps

Résumé du poste

Un institut de recherche en informatique propose un stage pour étudier l'utilisation des graphes de connaissances comme mémoire externe pour des agents conversationnels basés sur des LLM. Le stagiaire travaillera sur l'analyse des limitations et le prototypage d'une mémoire KG, tout en développant ses compétences en Python et en apprentissage automatique. Ce stage se déroule dans un environnement collaboratif à Sophia Antipolis, avec un focus sur la recherche innovante.

Prestations

Repas subventionnés

Remboursement partiel des frais de transport

7 semaines de congés annuels + 10 jours RTT

Télétravail possible après 6 mois

Équipement professionnel fourni

Accès à des événements sociaux et culturels

Formation professionnelle

Couverture sociale

Qualifications

• Étudiant(e) en Master 2 ou dernière année d'école d'ingénieur en informatique ou mathématiques appliquées.
• Connaissances en Python et apprentissage profond requises.
• Être curieux(se) et désireux(se) d'apprendre.

Responsabilités

• Étudier l'utilisation des graphes de connaissances comme mémoire externe.
• Analyser les limitations des agents basés sur LLM dans un scénario d'entreprise.
• Prototyper une mémoire KG pour la collaboration multi-agents.

Connaissances

Programmation Python

Machine Learning / Deep Learning

Systèmes multi-agents

LangChain

Web sémantique

RDF, RDFS, OWL

Formation

Master 2 ou école d'ingénieur

Outils

PyTorch

TensorFlow

Description du poste

Contexte et atouts du poste

The emergence of Large Language Models (LLMs) has recently accelerated the use and advanced integration of Artificial Intelligence in business tasks, most recently through conversational multi-agent systems. However, extended interactions between agents raise several continuity and consistency issues: loss of task context, history, or decisions, or exchange of redundant or contradictory information. These issues limit the use of LLM-based multi-agent systems in business tasks such as project management. Their mitigation is therefore an active research direction, for example with the design of an external memory [5,6]. In parallel, knowledge graphs (KGs) of the Semantic Web have been mentioned as a source of knowledge to complement LLMs and mitigate their hallucinations [3,4]. In particular, facts from KGs can be used to ground LLMs with processes such as Retrieval Augmented Generation (RAG) [1] or GraphRAG [2]. Interestingly, KGs could also be seen as an external memory for LLM-based agents, where facts could represent decisions, actions, and context. Such a representation could leverage existing ontologies such as PROV-O, Activity Streams, or FOAF. This line of research is associated with major challenges such as:

• The need to represent agents discussions, actions, decisions, results within KGs, potentially with different granularity levels
• The need to retrieve relevant context, actions, and results from KGs at the correct granularity level to support agents when they start a new task or encounter a blocking issue (e.g., contradictory information, loss of context)
• The need to detect those blocking situations

Mission confiée

In this internship, we propose to study the use of knowledge graphs as an external memory for a system constituted by LLM-based conversational agents.

This internship is a collaboration between the Wimmics team (Université Côte d'Azur, Inria, CNRS, I3S) and the Forgeron3 company. It will take place on the premises of the Wimmics team in Sophia Antipolis, in collaboration with Forgeron3 and under the supervision of:

• Pierre Monnin
• Fabien Gandon

Wimmics (Web-Instrumented huMan-Machine Interactions, Communities and Semantics) is a joint research team at Université Côte d’Azur, Inria, CNRS, I3S, whose research lies at the intersection of artificial intelligence and the Web. Wimmics members work on methods to extract, control, query, validate, infer, explain and interact with knowledge.

Forgeron3 develops a platform of collaborative intelligent assistants, based on open source LLMs such as those of Meta and Mistral. Forgeron3's goal is to democratize AI for European SMEs, allowing employees to focus on what matters while repetitive tasks are handled by intelligent assistants, improving every human interaction.

Principales activités

In this internship, we propose to study the use of knowledge graphs as an external memory for a system constituted by LLM-based conversational agents. In particular, the internship will include the following tasks:

• State of the art and skills development on LLMs, RAG, GraphRAG, Semantic Web, agents collaboration and memory
• Study of the limitations of an LLM-based agent collaboration from a company-based scenario
• Prototyping a KG memory for multi-agent collaboration
• Designing the KG: key entities, classes, relations, potentially re-using and adapting existing ontologies
• Designing a KG construction and completion process where agents complete the KG with relevant information
• Designing a retrieval process to enhance agents context when needed
• Experiment and evaluation of results.
• Definition of metrics of interest (e.g., information coherence, process achievement, performance of agents)
• Validation on a company-based scenario

Compétences

You are studying in Master Year 2 / final year of engineering school, with a specialty in computer science or applied mathematics. You are proficient in:

• Python programming
• Machine Learning / Deep Learning, especially with frameworks like PyTorch or Tensorflow
• Knowledge of LLMs, multi-agents systems, frameworks like LangChain, and (Graph)RAG would be appreciated.
• Knowledge of the Semantic Web (RDF, RDFS, OWL, SPARQL, knowledge graphs and ontologies) would be appreciated.
• Ability to read and write in English

You are curious, eager to learn, face challenges, experiment and discover by yourself.

Avantages

• Subsidized meals
• Partial reimbursement of public transport costs
• Leave: 7 weeks of annual leave + 10 extra days off due to RTT (statutory reduction in working hours) + possibility of exceptional leave (sick children, moving home, etc.)
• Possibility of teleworking (after 6 months of employment) and flexible organization of working hours
• Professional equipment available (videoconferencing, loan of computer equipment, etc.)
• Social, cultural and sports events and activities
• Access to vocational training
• Social security coverage