Highlights:

  • Generative AI models, like GPT-4, which runs ChatGPT, measure the similarity between vectors to build probabilistically constructed sentences or images based on prompts.
  • According to the company, niche databases for vector data are operationally complex and require developers to maintain a separate database, requiring teams to duplicate, synchronize, and track data across multiple systems.

Timescale Inc., the creator of a cloud time-series database that’s based on PostgreSQL, is seeking to cater to artificial intelligence developers with the introduction of its latest vector capabilities.

Timescale Vector, which was announced recently, resides on the company’s cloud-based PostgreSQL platform and enables the administration of vector embeddings alongside relational, time-series, analytics, and event-based data. The company promises that developers now have access to all the data required to fuel the most sophisticated AI models from a single database.

Timescale, which has raised more than USD 70 million in venture capital through a series of funding rounds, is the creator of a time-series database that records information in the chronological sequence in which it was generated. This is helpful for applications such as application logs, which characterize the latency experienced by computing systems over specific time periods.

It is also useful for financial applications, providing a means to track the fluctuating value of equities and shares. Utilizing time-series data, industrial companies monitor sensor measurements, such as the temperature of a piece of apparatus.

The database was originally developed as an initiative at the University of California, Berkeley, using the open-source PostgreSQL database. It is currently one of the most popular database formats in the world and is compatible with a variety of database management tools. So that developers will feel immediately at home when using the Timescale database.

With the addition of vector search capabilities, which is available for early access as of today, Timescale is broadening its horizons and positioning itself as a database capable of feeding next-generation AI applications. Vectors allow the storage of unstructured data such as images, audio, and handwritten text that cannot be represented by conventional databases. They can be viewed as the geometric or numerical representations of data with similar semantics.

Generative AI models such as GPT-4, which enables ChatGPT, function by measuring the similarity between vectors in order to generate sentences or images based on queries in a probabilistic manner. Consequently, AI models that can access vector-stored information can be considerably more potent than those that cannot.

According to Timescale, AI developers presently have two options. Either they can use a difficult-to-manipulate niche vector database, or they can use a more common database and add extensions for vector support.

According to the company, niche databases for vector data are operationally complex and require developers to maintain a separate database, requiring teams to duplicate, synchronize, and monitor data across multiple systems. In addition, engineering teams confront a precipitous learning curve when acquiring a new query language, system internals, application programming interfaces, and optimization techniques, according to the report.

According to Timescale, the inclusion of native vector capabilities simplifies the AI application architecture by requiring only a single database. In addition, since the Timescale database is essentially PostgreSQL with enhanced features, it inherits over 30 years of battle-tested robustness and dependability, according to the company.

In terms of advantages, vector capabilities will allow developers to speed up search with an innovative Approximate Nearest Neighbor index. According to the company, Timescale combines its ANN index with its own indexing algorithms to achieve 243% faster search speed than the specialist vector database Weaviate, and between 39% and 363% faster search performance than alternative PostgreSQL-based search indexes.

The time-series nature of Timescale also means it’s uniquely able to optimize time-based vector search to locate the most recent embeddings. It can additionally restrict vector search based on a time-based range or by the age of data, allowing it to retrieve historical data such as conversation histories.

Ajay Kulkarni, who helped start Timescale and is now its CEO, said, “The launch of Timescale Vector signifies our commitment to continuing to solve the biggest developer pain points so they can focus on building new AI applications more efficiently on a database foundation that’s fast, reliable and battle-tested.”

At launch, Timescale is collaborating with the AI application frameworks LangChain and LlamaIndex to integrate its database with two of the most widely used developer tools for creating AI applications.