Case Study
Objectives
Complex Administrative Documentation Navigation:
Cassandre is tailored to interpret and provide guidance on the multifaceted rules and documentation specific to different personnel categories within the Ministry of Education.
Accuracy and Factuality:
Ensuring that responses are not only rapid but also precise and grounded in factual information is a primary goal.
Rapidity and Comprehensiveness:
The tool aims to deliver comprehensive support operating much faster than manual or even embeddings-based research, improving the overall efficiency of HR services.
Challenges and Solutions
Complex Unstructured and Semi-Structured Documentation in Multiple Formats (pdf, text, images):
The diverse and complex nature of HR documentation within the Ministry of Education posed a significant challenge.
Ensuring Explainability:
To maintain the trust and confidence of HR agents, it is crucial to develop a system that not only provides accurate answers but also explains the reasoning and provides references leading to them.
Training and Adaptation:
Recognizing the importance of effective interaction between HR agents and the AI, tailored training programs were implemented to familiarize users with Cassandre's capabilities and optimize its usage in daily operations.
Technology and Methodology
Cassandre is a fine-tuned open weights (Mistral OpenHermès 7b) model which utilizes an interpretable Retrieval-Augmented Generation (RAG) infrastructure, integrating Knowledge Graph retrieval to enhance its ability to provide accurate and contextual answers.
The use of Knowledge Graphs which were processed and generated fully with pleias’ custom LLM largely improves the relevancy and precision of Cassandre’s responses, specifically in the cases where small employee groups with specific conditions are concerned.
