Rockset Inc., a real-time analytics database platform built for the cloud, today announced native support for hybrid search that incorporates vector search and text search along with metadata filtering into a single query.

As artificial intelligence models continue to grow in popularity the need for powerful data search and retrieval systems to augment them has grown alongside them. AI models, including Meta Platform Inc.’s Llama-3, OpenAI’s GPT-4, Google LLC’s Gemini, have led to a surge in applications that need access to both keyword search and vector search and the ability to index and rank vast amounts of data.

“AI models are getting better at an astounding pace,” Venkat Venkataramani, chief executive of Rockset, told SiliconANGLE in an interview. “But the models all have one real problem in that they don’t have any memory or knowledge, they can’t remember anything and they can’t memorize anything.”

To get that knowledge into AI models, engineers cobble it together from multiple sources, including vector databases, which index semantic data by its contextual “closeness” to other information, allowing information to be searched by how similar it is to other data. They also use traditional databases for keyword searches across structured data to pull back additional information. After that’s pulled in, they filter it again with metadata or geospatial data to narrow it down.

“There’s a lot of problems when you have multiple disparate systems, each doing a different part of the search,” said Venkataramani. “And you have to now combine them, the quality of your application suffers, the performance is lower and your application is not snappy.”

Venkataramani would know something about data because he worked for Facebook before it became Meta Platforms Inc., managing online data infrastructure as an engineering director. During his tenure, he helped Facebook scale its user-facing applications from around 30 to 40 million users to over a billion and a half until he left the company in 2015.

This is where Rockset’s hybrid search update comes in. It provides a hybrid search capability that does everything in one query, permitting developers to call on a vector search, a keyword search, metadata filtering and geospatial data all at the same time. There is no “duct taping” of multiple systems, as Venkataramani calls it, so everything happens all at once.

Rockset can even apply indexed ranking systems at the same time, delivering a powerful response for AI applications to understand the data that’s being retrieved.

“Because we’re built for the cloud, and a cloud-native database, there’s nothing to download, nothing to install, nothing to configure,” added Venkataramani.

Numerous applications could benefit from this sort of hybrid search. For example, an AI chatbot app for a real estate customer service company needs fast responses for customers looking for houses on the market. The app will need to reply with information on houses in the region the customer is in — quickly using geospatial metadata filtering as well as context-relevant dynamic ranking based on their preferences on what kind of houses they want to see and keywords related to types of houses.

“Without that retrieval augmentation, you can’t do anything to unlock the true power of these models and you discover every search becomes a hybrid search,” said Venkataramani. “Every large language model app and every AI app becomes a search application. There is no such thing as pure vector search, pure keyword search or pure geo search.”

Image: Shutterstock/Nepool