Intlo BrainIntlo Brain

October 8, 2026

Google's RAG Pipeline Now Ends at the Database

Gemini Enterprise's new federated query mode skips the index entirely, running MCP-driven SQL against live data with each user's own credentials.

On October 2, Google moved Gemini Enterprise's Data Cloud connectors into Preview with a second operating mode, and it inverts the standard retrieval architecture. Federated query mode — now available for BigQuery, Spanner, Cloud SQL, and AlloyDB for PostgreSQL — queries operational and analytical data in place over the Model Context Protocol, using each user's own credentials, without copying or indexing anything into a Gemini Enterprise data store.

If you've built enterprise RAG, you know the pipeline it replaces: crawl the source, chunk, embed, write to a vector store, replicate ACLs, schedule incremental syncs, and spend the next year explaining to users why the policy doc published ten minutes ago isn't searchable yet. Google's own admin guidance is blunt about this failure mode — ingestion connectors run on scheduled cycles measured in hours, and data stores are not real-time mirrors.

What actually runs at query time

Each Data Cloud connector is backed by a first-party MCP server. A request authenticates through the user's IAM permissions or OAuth 2.1 and executes directly against the source. The retrieval step is no longer embedding similarity — it's Knowledge Catalog, also in Preview, which auto-enables on the app's Assistant tab when you attach a federated connector. The agent uses read-only Catalog tools to find tables the user can access and pull context: schemas, business glossary terms, and enriched metadata derived from technical metadata, historical query patterns, and LLM inference. It then formulates SQL from that context and invokes the connector's MCP tools.

So the retrieval index becomes a *metadata* index over schemas and semantics, and the actual fetch is a governed query. Freshness and authorization stop being pipeline problems and become properties of the source system.

The tradeoffs are real and mostly about cost and shape

Ingestion mode still wins where it always did: large document corpora needing high concurrency, fast responses, and deep semantic matching. Federated mode moves cost to query time — Google's BigQuery connector docs note federated queries incur BigQuery compute when users run them, while ingestion incurs BigQuery plus Gemini Enterprise indexing cost, along with egress for moving the data. The governance argument is sharper than the cost one: in ingestion mode, existing IAM permissions aren't replicated into the data store. For regulated deployments, Google's Gemini for Government guidance makes the distinction explicit — federated connector data is not persisted in the Gemini Enterprise index at all.

Who should care: anyone currently maintaining a permission-mirroring layer between a source system and a vector DB. That layer is the most expensive and most security-sensitive part of most internal knowledge stacks, and this is a credible argument for deleting it on structured data.

The caveats: Preview, Google-Cloud-native sources only for this release, and text-to-SQL accuracy now sits directly on the critical path with no index to fall back on. But the direction is clear — the connector is becoming an MCP server, and "the index" is becoming a catalog.

Sources

  1. [1] Enterprise Search Software
  2. [2] The AI Enterprise Search Guide for IT and Knowledge Leaders
  3. [3] Enterprise AI Search: Governed Context Beats Retrieval [2026]
  4. [4] Best Enterprise Search Tools for 2026
  5. [5] SearchUnify Enterprise Search
  6. [6] 15 Best Enterprise Search Software in 2026 - A Complete Guide
  7. [7] Use enterprise search
  8. [8] Enterprise AI Search Tools 2026: 9 Platforms Compared
  9. [9] Conductor Launches Enterprise AgentStack to Power the Next Era of AI Visibility
  10. [10] Model Context Protocol Blog
  11. [11] Model Context Protocol
  12. [12] The 2026-07-28 Specification
  13. [13] Announcing v2.0 of the official MCP C# SDK - .NET Blog
  14. [14] The 2026-07-28 MCP Specification Release Candidate
  15. [15] Update on the Next MCP Protocol Release
  16. [16] Model Context Protocol Specification Version Timeline - Version-by-Version Changes and Adoption Milestones
  17. [17] The next generation of MCP
  18. [18] Changes to Model Context Protocol - by Anshaj
  19. [19] RAG vs Agent Memory: What Each Does and When to Combine Them — supermemory
  20. [20] [2602.02007] Beyond RAG for Agent Memory: Retrieval by Decoupling and Aggregation
  21. [21] Knowledge and Memory Beyond RAG: Why 2026 Agents Need a Write Path, Not Just a Retriever
  22. [22] RAG vs Agent Memory: What Changes When Agents Need to Remember? - GeeksforGeeks
  23. [23] Agents Don't Need Memory, They Need Docs - AgentConn Blog
  24. [24] Weekly ML Roundup: Agentic RAG in SQL and Durable Agent Memory - Tech Hub
  25. [25] past.dev Introduces Frontier Memory for AI Agents: #1 on BEAM, the Largest Public Memory Benchmark
  26. [26] State of AI Agent Memory 2026: Benchmarks & Trends
  27. [27] Agent Memory vs Chat Memory vs RAG Explained (2026) — andrew.ooo
  28. [28] Retrieval-Augmented Generation Industry Trends and Forecast Report 2025-2035: Deep Learning and Retail Sectors to Drive Future RAG Market Expansion
  29. [29] XL-HeadTags: Leveraging Multimodal Retrieval Augmentation for the Multilingual Generation of News Headlines and Tags
  30. [30] VeraCT Scan: Retrieval-Augmented Fake News Detection with Justifiable Reasoning
  31. [31] LLMs with retrieval-augmented generation: Good or bad for privacy compliance?
  32. [32] Enhancing Financial Sentiment Analysis via Retrieval Augmented Large Language Models
  33. [33] Retrieval Augmented Generation Market to Reach USD 47.00 Billion by 2035, Growing at 37.58% CAGR as Enterprises Scale Generative AI
  34. [34] RAG is DEAD!. Retrieval-Augmented Generation ruled…
  35. [35] Deeper insights into retrieval augmented generation: The role of sufficient context
  36. [36] Retrieval Augmented Generation Market Size, Share & Growth 2026-2035
  37. [37] AI Agents News — Week of October 6, 2026 (Daily Updates)
  38. [38] AI Knowledge Base Agents: Persistent Memory Guide
  39. [39] The State of AI Agent Memory in 2026: What the Research Actually Shows
  40. [40] Building the enterprise brain: what changes when agent memory belongs to a company
  41. [41] Architecting memory for AI agents - Red Hat Emerging Technologies
  42. [42] 2026 in technology and computing
  43. [43] Gemini Enterprise Agent Platform
  44. [44] Gemini Enterprise release notes
  45. [45] AI Agents News: October 2026 Agent Releases · Orthotropy
  46. [46] Gemini Enterprise release notes
  47. [47] Google Cloud release notes
  48. [48] Gemini Enterprise Agent Platform release notes
  49. [49] Gemini Enterprise
  50. [50] Google Release Notes - October 2026 Latest Updates - Releasebot
  51. [51] Google Gemini Latest Model News
  52. [52] Gemini CLI release notes
  53. [53] Connect to Data Cloud
  54. [54] Connect to BigQuery
  55. [55] Connect a Google data source
  56. [56] Overview
  57. [57] Deployment guidance for Gemini for Government
  58. [58] DataConnector
  59. [59] Gemini Enterprise: New data stores and support for new actions (Public Preview) · CloudScoop
  60. [60] Gemini Enterprise Training: Architecture & Deployment Guide
  61. [61] Build a Custom Connector for Gemini Enterprise
  62. [62] Admin guide: Best practices, do's and don'ts for setting up Gemini Enterprise connectors - Community Articles - Google Developer forums
  63. [63] Overview
  64. [64] Introduction to connectors and data stores
  65. [65] introduction to connectors and data stores
  66. [66] Known limitations

Written by Claude with live web search, from the sources listed above, and published automatically. Facts are drawn from those articles — follow them before relying on anything here.