From the course: LLM Foundations: Vector Databases for Caching and Retrieval Augmented Generation (RAG)

Pros and cons of vector databases

From the course: LLM Foundations: Vector Databases for Caching and Retrieval Augmented Generation (RAG)

Pros and cons of vector databases

“

In this video, we will discuss some advantages and shortcomings of vector databases, especially the specialized offerings. Let's begin with the advantages. Vector databases support semantic search. They have built-in implementations of approximate nearest neighbor algorithms, as well as a few distance measures. They support bulk data loading, which helps quickly load up large chunks of data like documents. They have indexing for vectors. This helps create indexes on vector fields and helps in executing semantic searches with low latency. They do have efficient data retrieval methods, especially for the large vector stores. They can scale well and can support high data and query volumes. Clustering and fault tolerance capabilities are also available to help with scale and redundancy. They are built for critical production applications. What are some of the shortcomings? They have limited support for traditional querying. Popular RDBMSs support several capabilities like joints and aggregations, but they are limited in vector databases. Similarly, the support for built-in functions like date and string manipulations are also limited. Transactional support is limited when it comes to achieving high levels of asset support. They do exhibit latency when it comes to ingesting large data sets in a short amount of time due to index processing. Semantic searches are computationally expensive and may sometimes require GPUs for low latency. They are also memory-intensive as indexes need to be reloaded into memory for searching purposes. Unlike traditional RDBMS, integrations into third-party products and tools are also limited. Vector databases are a fast growing domain triggered by the explosion of LLMs. We can expect that these limitations will be overcome with newer versions of these products over time.

Contents