StarTree boosts AI agent support in its real-time analytics platform

StarTree Inc., the developer of a managed service based on the Apache Pinot real-time data analytics platform, is enhancing support for artificial intelligence workloads with two enhancements being announced today.

They include support for Anthropic PBC’s Model Context Protocol and vector embedding model hosting. MCP provides a standardized way for AI applications to connect with and interact with external data sources and tools to extend their built-in knowledge.

Vector embedding model hosting allows machine learning models to convert multimodal data types such as text, images and audio into dense numerical representations that capture the semantic meaning of the input so it can be accessed via an application program interface or integrated directly into applications. That allows for advanced pattern matching and similarity searches on the data, going beyond text matching.

These combined capabilities enable StarTree to support agentic AI applications, real-time retrieval-augmented generation or RAG, and conversational querying of real-time data.

MCP is intended to spare developers the hassle of writing large amounts of custom code to integrate outside sources. While currently aimed mainly at developers, “MCP has the potential to benefit almost every stakeholder in the world of AI,”  according to Jason Andersen, vice president and principal analyst at Moor Insights & Strategy. It functions as an application program interface that eliminates the need for developers to each build bespoke integration hubs.

Real-time agents

MCP support allows AI agents to dynamically analyze live, structured enterprise data from StarTree’s high-concurrency architecture, simplifying the deployment and management of autonomous agents. It also makes natural language-to-SQL queries easier and less brittle to deploy and enables conversational questions to build upon previous answers, StarTree said.

“The MCP server allows AI agents to retrieve contextual information in a scalable manner,” said Chinmay Soman, head of product. “It enriches every decision that the AI agent makes with fresh data and handles thousands of concurrent queries per second.”

Agents can also use the server to search for services that satisfy specific requests and connect to them directly. “It can discover schemas and interesting data or insights automatically by essentially having a conversation with the database through the MCP server,” Soman said.

Vector embeddings allow queries against data types that don’t lend themselves well to conventional SQL. StarTree’s new vector auto-embedding enables pluggable vector embedding models to streamline the continuous flow of data from source to embedding creation to ingestion. This enables RAG to be done in real time for uses like financial market monitoring and information technology infrastructure observability.

RAG time

“Traditional RAG is pretty batch-oriented,” Soman said. “In a case like stock trading, prices can move based on comments on TV or stock filings. You can ingest that data into Pinot and ask questions like how stock is likely to trade that afternoon based on the freshest information.”

Pinot has supported vector embedding for over a year, but “we are doing it natively in the database so if something changes, we automatically reflect the latest embedding for a given record,” Soman said. For example, observability log data can be translated into embeddings and searched immediately. “You can actually have a conversation with your logs,” he said.

“If you have exact pattern-matching, then text indexing works fine,” said Peter Corless, director of product marketing. “But if you want to see if one log incident is like another log incident, you need vector similarity search as well as text indexing. StarTree now provides that capability natively, whereas other analytical databases require one database for vectors and another database for text indexing, he said.

StarTree also announced the general availability of Bring Your Own Kubernetes, a new deployment option that gives organizations full control over StarTree infrastructure within their own Kubernetes environments, whether in the cloud, on-premises or in a hybrid architecture.

This model is targeted at regulated industries where data residency, compliance and security policies limit cloud processing. It’s also a more cost-effective option for organizations with stable, predictable workloads because it saves on computing and egress fees, StarTree said.

The company previously offered software-as-a-service and “bring your own cloud” options, but the latter requires delegated access into a customer’s cloud account. “That’s OK for most customers but for some it’s a point of friction,” Soman said. “This model is completely disconnected; we don’t have any connection to the data plane whatsoever.”

MCP support will be available in June, with vector embedding due to arrive in the fall.

Photo: Unsplash