Timescale Inc., a PostgreSQL cloud database company, today announced the release of two new open-source extensions that significantly enhance its scalability and ease of use for artificial intelligence applications when retrieving data from vector databases.

The new extensions, pgvectorscale and pgai, bring to bear the ability to use PostgreSQL, an open-source relational database for vector data retrieval, which is fundamental for building AI applications and specialized contextual search.

Vector databases allow AI developers to insert data into high-dimensional arrays so that they are connected by how they are related to one another contextually. Unlike standard relational databases, vector databases store data using contextualized meanings where the “nearest neighbor” can be used to connect them, such as how a cat and a dog are closer in meaning (as household pets) than a cat and an apple. This makes it faster to find information when an AI performs searches for semantic data such as keywords, documents, images and other media.

Most of this data is stored in extremely popular, high-performance vector databases, Avthar Sewrathan, AI product lead at Timescale, told SiliconANGLE in an interview, however, that not all the data that services use is stored in vector databases. As a result, sometimes multiple data sources exist in the same environment.

“Everyone at every organization in the world is trying to integrate AI in some way, whether it’s building new applications that take advantage of large language model capabilities, or whether it’s reimagining existing applications,” said Sewrathan. “So, when CTOs or engineering teams go and figure out how we can do with AI, they face this choice do you use a separate vector database, or do you use a database that you already know and that you’re familiar with? The motivation behind these extensions, simply put, is making Postgres a better database for AI.”

The first extension, pgvectorscale, builds on an already existing open-source extension pgvector, and helps developers build more scalable AI applications with higher performance embedding for search at lower cost.

Sewrathan said it includes two innovations, DiskANN, a capability that is capable of offloading half of its search indexes to disk, which are normally stored in memory with very little hit to its performance, which helps save greatly on costs and Statistical Binary Quantization, an improvement on standard binary quantization that helps reduce memory use.

Timescale’s benchmarks of pgvectorscale revealed that PostgreSQL could achieve 28x lower latency for 95% and 16x higher query throughput compared to the popular Pinecone vector database for approximate nearest neighbor queries at 99% recall. In contrast to pgvector, which is written in C, pgvectorscale is written in the Rust programming language, which will open up more opportunities for PostgreSQL developers to build for vector support.

The next extension, pgai, is designed to make it easier for developers to build search and retrieval-augmented generation, or RAG, solutions for their AI applications. RAG combines the strengths of vector databases with the capabilities of LLMs by providing them with real-time, up-to-date authoritative information to help reduce the incidence of hallucinations, or when an AI confidently issues false statements.

The technique is fundamental to building accurate and reliable AI applications. The initial release of pgai supports creating OpenAI embeddings quickly, and the release is building OpenAI chat completions from models such as GPT-4o directly within PostgreSQL.

OpenAI’s GPT-4o is OpenAI’s newest flagship model that provides powerful multimodal capabilities that include real-time voice conversation and video understanding.

Sewrathan said that by having the ability to use vectors in PostgreSQL creates a powerful “ease of use” bridge for many developers this can be very powerful as many businesses already rely on a PostgreSQL database or some other relational database.

“So that ability to then add on vector storage and such capabilities via an extension is a lot easier because it simplifies your data architecture,” Sewrathan said. “You have one database. It can store multiple data types in the same place. That’s been a huge thing, because otherwise, what you have is this data synchronization, data deduplication and a lot of complexity.”

Image: SiliconANGLE/Microsoft Designer