School of Engineering Department of Computer Science and Engineering 136 Retrieval Augmented Generation with Vector Database Supervisor: ZHOU Xiaofang / CSE Student: CHOW Wang Hin / COMP Course: UROP 1100, Spring Large language models have achieved remarkable success in generating coherent and fluent text, but they often suffer from hallucinations, inaccuracy, and a lack of factual consistency, which may lead to misleading and unreliable responses. This limitation is particularly worrisome in applications where accuracy and trustworthiness are critical, such as question answering and conversational AI. To address these limitations, retrieval augmented generation has emerged as a promising approach. By retrieving relevant information from vector databases and incorporating it into the generated text, the models can provide more accurate and informative responses. This report presents a comprehensive and in-depth exploratory study in retrieval augmented generation, focusing on sophisticated and robust multimodal embedding techniques. Retrieval Augmented Generation with Vector Database Supervisor: ZHOU Xiaofang / CSE Student: FU Yixuan / COSC Course: UROP 1100, Summer Mixed-modality data is increasingly being utilized in the modern industry, whose similarity search is often accompanied by predicate filtering. To better leverage these data, there is a need for a method that can efficiently and accurately establish an index and conduct similarity searches with structured predicates. To address this issue, four scholars have proposed ACORN, a novel and efficient method for similarity search that supports predicate query and filtering. It is based on Hierarchical Navigable Small Worlds (HNSW) and proposes a new concept of predicate subgraph traversal. The scholars evaluated the performance of ACORN across various datasets, and the results demonstrate its state-of-the-art capabilities on all datasets. Retrieval Augmented Generation with Vector Database Supervisor: ZHOU Xiaofang / CSE Student: LI Xinwei / COGBM Course: UROP 1100, Summer With rapid advancements in deep learning models, vector embeddings are now a crucial data format used in various applications such as retrieval-augmented generation and similarity search. This has led to an increased adoption of vector databases and indices, providing an effective way to conduct approximatenearest-neighbor (ANN) searches across extensive data. However, in a hybrid search setting, integration of ANN similarity search and predicate filter is becoming more and more necessary, yet current predicate filtering algorithms often come with restricted search predicates or poor performance. To address this, ACORN, a predicate-agnostic algorithm, has been implemented. This report will introduce ACORN's search and construction algorithms, evaluate its accuracy and efficiency, and implement some Milvus and FAISS codes for later reproduction of ACORN.
RkJQdWJsaXNoZXIy NDk5Njg=