Agentic AI
Pinecone Nexus Lets AI Agents Query Structured Business Data Directly
The new knowledge engine transforms scattered enterprise documents into queryable structured data, helping AI agents work more efficiently with less token overhead.
InfoQ
Pinecone has launched Nexus, a knowledge engine that bridges the gap between raw enterprise data and what AI agents actually need to work effectively. Rather than forcing agents to search through unstructured documents on every task—burning tokens and slowing performance—Nexus lets organizations compile their business context once into a structured layer agents can query directly.
The Problem It Solves#
Large language models excel at general knowledge, and vector databases are good at finding buried information. But enterprises need something different: access to business context that typically sits scattered across contracts, wikis, HR documents, meeting notes, support tickets, and financial records. When agents must retrieve and synthesize this information during every task, the result is inefficient query loops, inflated token costs, and incomplete answers.
How Nexus Works#
The system organizes knowledge around three key concepts: workspaces (top-level containers tied to teams or business units), contexts (specific datasets or knowledge domains), and manifests (blueprints for ingesting raw data and converting it into structured knowledge). Manifests are where domain expertise enters the system—subject matter experts can design the artifact types and relationships that encode their knowledge before any agent query runs, letting agents inherit that understanding rather than reverse-engineering it at runtime.
Data flows in through connectors supporting local files, Box, and Microsoft OneLake, with Google Drive, Slack, GitHub, Notion, Confluence, and S3 coming soon. Once ingested and curated, agents and chatbots query the data using KnowQL. Nexus also includes a preview playground for testing data connections and validating approaches, plus BYOC (Bring Your Own Cloud) deployment for organizations with strict data residency and compliance requirements.
Early Results#
Early adopters in financial services and legal research have reported striking improvements. In legal tasks, Nexus achieved 100% completion versus 6% for a coding agent and 66% for a traditional RAG system. The RAG system faltered on doctrine synthesis, cross-case reasoning, and multi-source questions. Token spend dropped 9–15× over conventional retrieval. Enterprise data management tasks showed 90% accuracy compared to 65% for RAG, with curation costs around $0.0038 per document.
