Qdrant launches Cloud Inference for multimodal vector search with higher speed and lower cost – News

Qdrant, the leading provider of high-performance, open-source vector search, today announced the launch of Qdrant Cloud Inference, a fully managed service that enables developers to search both text and image vectors simultaneously.

Normally, developers must create artificial intelligence workflows that require separately building and embedding information about unstructured documents and images, distinct from the images themselves. This means that when images and text are uploaded together, such as in a document that contains both, they must be separated and split into distinct pipelines, creating a fragmented flow.

“Traditionally, embedding generation and vector search have been handled separately in developer workflows,” said André Zayarni, chief executive and co-founder of Qdrant. “With Qdrant Cloud Inference, it feels like a single tool: one [application programming interface] call with optimal resources for each component.”

According to Qdrant, this allows for a much more elegant approach with a solution that allows developers to combine dense and sparse unstructured documents alongside images in one environment, helps reduce latency, cuts network costs and simplifies search.

Vector databases are fundamental for providing essential memory for AI assistants and agents, which can allow them to retain context over extended periods of time. They essentially act as a kind of semantic memory for AI; they help assistants and agents “remember” and retrieve information based on meaning separated from the documents and images uploaded by businesses and providers.

Having an infrastructure that can update vector databases quickly means that information in them can be made up-to-date with lower overhead. Qdrant also believes that its service will help reduce the total number of times developers need to make API calls, driving less cloud usage and AI models, meaning that it will lower costs.

“Embedding generation and management is often a fragmented, complicated workflow for developers building AI-driven applications,” said Paul Nashawaty, practice lead and principal analyst for theCUBE Research. “Seventy-five percent of organizations are using 6 to 15 tools for management.”

When a user asks an AI chatbot a question or an AI agent begins to work through a problem, it hooks into a broad variety of sources and search is at the ground level. Oftentimes, AI agents, especially research agents, run through tens, if not hundreds, of searches to begin progress towards a goal. As a result, milliseconds can matter depending on the search infrastructure.

According to Qdrant, its setup uses separate image and text embedding models natively integrated into its cloud service to build vector databases. Vector embeddings are numerical representations of data, in this case, that can include words, images or even entire documents, that capture the semantic meaning and their relationships. These numbers transform the complex data into a format that machine learning and AI models can understand and process into similarity searches.

When both images and text are submitted together, both the text and images are processed at the same time and embedded into the vector database, permitting them to be searched at the same time using a multimodal search capability. The supported models and algorithms include MiniLM, SPLADE, BM25, Mixedbread Embed-Large and CLIP for both text and image.

Qdrant said the new paid offering includes up to 5 million free tokens per month, with unlimited tokens for BM25, which is a term-based ranking algorithm for text search often used alongside vector embeddings for hybrid semantic retrieval. This will allow development teams to test, build and iterate on AI features such as multimodal search, retrieval-augmented generation and hybrid search.

Image: Nepool/Shutterstock