Utilizing Vector Databases for Similarity Search

This bachelor’s thesis explores applications of two vector database systems, ChromaDB and Qdrant, with focus on their comparison from a user’s perspective and the effectiveness of specific text representations and distance metrics in similarity search. In order to perform the comparison, three datasets are utilised, one from the aviation domain and two from research and development projects. The initial phase involves familiarising with TF-IDF representation, MiniLM-L6 and SPECTER models. Standard similarity metrics and evaluation metrics, rank-biased overlap, and precision at k, are introduced. The datasets are pre-processed and then used to fill the databases. The similarity search evaluation is done concerning humansorted documents based on their similarity to the query. Based on experiments with various representations and distance metrics, those created by the MiniLM-L6 model emerge as the best representation; when changing the distance metric, the results do not differ significantly.