Teradata adds agentic, multimodal AI to vector store

Teradata has added agent-based features and multimodal data support to its Enterprise Vector Store, expanding how organisations index and search unstructured content alongside traditional enterprise data across cloud and on-premises systems.

The update more tightly links Teradata's vector store with Unstructured, the data-processing software provider, and adds functions for building and running AI systems that work with text, images, and audio. Video support is planned.

Vector stores underpin many retrieval-augmented generation deployments, where AI models fetch relevant internal information before generating a response. Teradata positions Enterprise Vector Store as a unified layer for embeddings, indexing, and search across different operating environments.

Multimodal pipeline

Enterprise Vector Store now supports a workflow from embedding generation to indexing and metadata management, with integration into AI development frameworks.

Integration with Unstructured adds automated ingestion and processing of documents, PDFs, images, and audio. The release also adds multimodal embeddings for text, image, and audio, plus support for higher embedding dimensions up to 8K.

Teradata has also introduced hybrid search, combining semantic vector search with lexical search and metadata-driven techniques, aimed at improving retrieval accuracy and context from enterprise data.

The release also adds LangChain integration, providing direct connections for retrieval-augmented generation workflows and agent execution. Agent behaviour is intended to extend into governed actions and workflow orchestration.

Scaling pressures

Many organisations are seeing rapid growth in unstructured information such as documents, media files, and recorded audio. Teradata cited a Gartner estimate that unstructured data is growing at three times the rate of structured data as businesses store and analyse content beyond traditional tables and logs.

At the same time, more AI models accept multiple input types, increasing interest in systems that can search and retrieve across modalities rather than treating text, images, and audio as separate stores.

Teradata also cited survey-style research suggesting AI agent adoption is rising, saying nearly 80% of companies are already deploying AI agents, with many projecting returns from agentic AI initiatives. Enterprises still report barriers including fragmented data silos, limits in scaling infrastructure, and difficulty providing unified access across structured and unstructured sources.

Performance claims

Teradata argues that scale remains a constraint in much of the vector database market. It referenced Forrester research noting that "high-end scale and performance still require considerable effort, especially when supporting tens of billions of data points (vectors)."

Teradata says many vector offerings hit practical limits at a few hundred million embeddings. By contrast, it says Enterprise Vector Store is designed to ingest millions of documents and handle thousands of files per hour, depending on configuration and data characteristics.

The company also says the product can scale across billions of vectors and higher-dimensional embeddings, and handle more than 1,000 concurrent queries without performance degradation. It also positions the system as a way to reduce duplicated infrastructure by consolidating vector workloads into existing data architecture, while maintaining governance controls across cloud, on-premises, and hybrid deployments.

Use cases

The update targets organisations building agent-driven workflows in which systems retrieve context and then take actions based on policy and audit controls. Teradata says the combination of Unstructured processing and LangChain integration is designed to simplify how unstructured content is prepared and accessed by agents.

The company highlighted healthcare scenarios that combine structured patient records with clinical notes, medical images, and audio dictations. It also pointed to insurance claims use cases, with agents processing damage photographs and policy documents alongside structured claims data and coverage rules.

Other examples included defence-oriented scenarios involving image capture from mobile applications and retrieval from a vector store alongside terrain patterns and other signatures. Teradata also described financial services teams using unstructured policy definitions and structured business data to answer questions about loyalty discount eligibility.

Executives framed the product changes as part of a broader shift from chat-style interfaces to agent-led systems that run workflows and enforce governance rules.

"We're entering an era where AI agents will become the primary interface for enterprise intelligence-autonomously orchestrating workflows, making decisions within defined governance frameworks, and uncovering insights across every data type," said Sumeet Arora, chief product officer at Teradata.

Arora added: "Stand-alone vector databases can't deliver on this vision. Teradata Enterprise Vector Store fundamentally reimagines how enterprises operationalise AI by unifying structured and multi-modal unstructured data with autonomous agent capabilities within a single governed platform. Organisations can now move from prototype to production-grade agentic systems in some cases within hours, not months-while maintaining the governance, security, and sovereignty that mission-critical AI demands."

Unstructured founder and CEO Brian Raymond focused on data handling and governance controls within the platform.

"Enterprises shouldn't have to choose between data security and AI readiness. By embedding Unstructured natively inside Teradata Enterprise Vector Store, Teradata customers get production-quality, AI-ready data at scale, with no external tools, no data leaving the platform, and no compromise on governance," said Raymond.

The new agentic and multimodal functions are set for general availability to Teradata customers from April 2026.