Compatibility with Multiple Embedding Models in Vector Databases

This thesis investigates the compatibility of sentence embedding across open-source embedding models by implementing and evaluating existing embedding conversion methods. The goal is to assess whether embedding by one model can be translated into the embedding space of another model without modifying the embedding model themselves. Using the open-source models, MiniLM, MPNet and DistilBert, the conversion method is applied across different model pairs and the resulting embeddings are evaluated through t-SNE visualizations and various retrieval metrics. Experiments show that direct retrieval between different models typically results in poor performance due to architectural differences, but that the converter improves compatibility across models. The converter also enables inter-operability between models with different embedding dimensionalities, allowing them to coexist within the same retrieval system. These results demonstrate that embedding conversion is a viable approach to enable cross-model interoperability and flexibility in embedding-based systems.