Agents Need Bundled Context Over Classic RAG Chunks

The Breakthrough

Pinecone demoted pure vector search with Nexus and NoQL, a retrieval contract that carries intent, filters, access policy, provenance, response shape, confidence, and budget to assemble agent bundles like customer context + entitlement + policy + history.

What Actually Worked

• Pinecone Nexus defines a retrieval contract for agents needing operating context: it assembles customer record, entitlement, controlling policy, and prior history into a usable bundle, rather than searching systems from scratch each run.
• PageIndex builds a hierarchical tree of documents like a table of contents with summaries on each node; the model reasons through the tree without embeddings or chunking to preserve structure, claiming 98.7% accuracy on FinanceBench for filings where sections like risk factors differ from management discussion.
• SAP acquired Dreamio for lakehouse architecture, semantic layer, query federation, access controls, and lineage across systems, plus Prior Labs' TabPFN tabular foundation model (published in Nature) for table-native reasoning over ERP/CRM data like revenue metrics or supplier risk.
• Microsoft GraphRAG handles relational knowledge like supplier-shipment connections or shared failure patterns via graphs, as chunks and tables cannot carry inherent relationships.
• Write the agent's data contract first: list the bundle (e.g., for refund agent: customer record, plan, region, product version, purchase history, refund policy, threshold, prior exceptions, current ticket, approved language, refund authority); then select primitives (vector for prose, trees for docs, semantic layer for tables, graphs for relations).

Before / After

Pinecone claims rediscovery eats up to 85% of agent compute. PageIndex hits 98.7% accuracy on FinanceBench evaluation versus chunk-based retrieval.

Context

Classic RAG suits chatbots answering from 3 similar chunks (e.g., password reset), but agents perform tasks like opening tickets, retrieving records, checking policies, and drafting responses, requiring bundled context assembled consistently without refetching or re-summarizing each run. Vendors race to provide knowledge layers matching data shapes: prose chunks, document hierarchies, governed tables, relational graphs. Bigger context windows fail due to context rot, lack of authority marking, permissions, hierarchy preservation, and reliable emphasis; Chroma research shows performance degrades in cluttered long contexts.

Notable Quotes

• "Pine Cone says that that kind of rediscovery can eat up to 85% of agent compute."
• "The retrieval unit needs to match the work you're doing... a chunk works for a simple FAQ, a section works for a filing, a table works for financial analysis."
• "Don't pick a database first. Pick the contract your agent will have with the data first."

Content References

Full article linked in description.