A PostgreSQL cloud database provider recently declared the availability of two brand-new, open-source extensions that greatly improve the scalability and usability of its data retrieval from vector databases for artificial intelligence applications.
Using PostgreSQL, an open-source relational database, for vector data retrieval is made possible by the new extensions, pgvectorscale and pgai. This is essential for developing AI applications and specialized contextual search.
AI programmers can add data to high-dimensional arrays using vector databases, connecting them based on their contextual relationships with each other. Vector databases store data using contextualized meanings, where the “nearest neighbor” can be used to connect them, in contrast to typical relational databases. For example, a cat and a dog have a closer meaning as family pets than does an apple. When an AI searches for semantic data, including keywords, documents, photos, and other media, this speeds up the information-finding process.
Timescale’s AI product lead, Avthar Sewrathan, told SiliconANGLE in an interview that while most of this data is kept in very popular, high-performance vector databases, not all of the data used by services is kept in vector databases. Thus, in the same context, there are occasionally several data sources.
“AI is being incorporated into every organization in the world, in some form or another, whether through the development of new apps that capitalize on the power of large language models or through the redesign of current ones,” stated Sewrathan. Therefore, CTOs and technical teams must decide whether to employ a distinct vector database or a database they are already familiar with while figuring out how to use AI. Encouraging Postgres to be a better database for AI is the driving force behind these enhancements.
Building on the open-source foundation of the original expansion, pgvectorscale, enables developers to create more scalable artificial intelligence (AI) applications with improved search performance at a reduced cost.
According to Sewrathan, it incorporates two innovations: Statistical Binary Quantization, which is an enhancement of standard binary quantization that helps reduce memory use, and DiskANN, which can offload half of its search indexes to disk with very little impact on performance. DiskANN is capable of saving a significant amount of money.
In comparison to the widely used Pinecone vector database, PostgreSQL was able to attain 28x lower latency for 95% and 16x greater query throughput for approximate nearest neighbor queries at 99% recall, according to Timescale’s benchmarks of pgvectorscale. Since pgvectorscale is written in Rust instead of C, PostgreSQL developers will have more options when developing for vector support.
The next addition, pgai, is intended to facilitate the development of retrieval-augmented generation, or RAG, solutions for search and retrieval in applications using artificial intelligence. In order to lessen the frequency of hallucinations—which occur when an AI boldly makes erroneous statements—RAG blends the advantages of vector databases with the skills of LLMs by giving them access to current, reliable information in real-time.
Building precise and dependable AI systems requires an understanding of this technique. OpenAI conversation completions from models like GPT-4o are built directly within PostgreSQL with the first release of pgai, which facilitates the creation of OpenAI embeddings rapidly.
The most recent flagship model from OpenAI, the GPT-4o, offers strong multimodal capabilities like video comprehension and real-time speech communication.
According to Sewrathan, PostgreSQL’s vector functionality builds a strong “ease of use” bridge for developers. This is significant because many firms currently use PostgreSQL or other relational databases.
Because it streamlines your data architecture, adding vector storage and other features via an extension is much easier, according to Sewrathan. “One database is all you have.” It has the ability to store several data kinds simultaneously. That has been extremely beneficial because without it, there would be a great deal of complexity, data synchronization, and data deduplication.