Soumia DAAS

As part of the Research & Innovation department (DRIT), I worked on improving Retrieval-Augmented Generation (RAG) systems by adapting semantic embeddings for domain-specific corpora. My mission focused on the fine-tuning and linear transformation of embedding spaces to enhance information retrieval performance in specialized datasets. I conducted research, experimentation, and evaluation of alignment methods to optimize similarity search and query-document matching. Key contributions: Designed and implemented a complete pipeline for embedding adaptation using PyTorch and OpenAI embeddings Generated synthetic queries and constructed positive/negative pairs for training Trained a linear transformation matrix (embedding adapter) with CosineEmbeddingLoss and triplet evaluation Indexed and evaluated the corpus using Weaviate vector database, measuring MRR and Hit@K metrics Documented all experiments and results on GitLab and Confluence for reproducibility Tools & Technologies: Python, PyTorch, OpenAI API, Weaviate, NumPy, scikit-learn, pandas,JSONL