Shared, managed retrieval
Foundry IQ is a managed knowledge platform for agents, built on Azure AI Search, that provides shared retrieval without building custom RAG infrastructure per agent. It solves the duplication problem: multiple agents share one set of knowledge bases, so you focus on agent experience rather than plumbing.
Key concept · Knowledge bases organized by domain
A knowledge base organizes information by business domain — "Product Documentation," "HR Policies" — not by technical storage location. One base can consolidate multiple sources: specs from SharePoint, API docs from Blob Storage, analytics from OneLake, tickets from existing indexes. Agents search a unified domain instead of "Site A" or "Container B," and one base can serve several agents at once (improvements benefit all of them immediately).
Connect once, retrieve intelligently
Foundry IQ connects to existing storage (SharePoint, Blob, OneLake) and handles the pipeline automatically: discovery (scan for documents) → processing (chunk and embed) → indexing (make searchable) → monitoring (reindex on change). Configure once per source; all connected agents benefit.
Key concept · Built-in retrieval intelligence
Foundry IQ applies retrieval strategies with no custom code: query analysis (understand the information need), strategy selection (keyword for simple facts, semantic + query expansion for complex questions), result ranking (relevance scoring surfaces the best content, reducing tokens), and citations (source references for trust and verification).
Agents reach a knowledge base through the Model Context Protocol — an MCPTool pointed at the knowledge base's MCP endpoint. The agent retrieves like any other tool, with no custom retrieval logic.
Grounded, cited, and safe when unknown
Well-indexed content still fails if the agent answers from training data or searches without citing. The fix is precise agent instructions that specify three behaviors: when to retrieve (always search, never rely on training data), how to cite (an exact citation format), and fallback (what to say when information isn't found).
| Behavior | Verdict |
|---|---|
| Answer from training data | ✕ Wrong / generic information |
| Search but don't cite | ✕ Unverifiable, no accountability |
| Search + cite + ground | ✓ Enterprise standard |
Tailor instructions to agent type (customer support: never guess, escalate; internal research: synthesize with confidence notes; compliance: cite specific policy sections), test across straightforward/synthesis/out-of-scope/ambiguous queries, and monitor citation and fallback frequency in production.
Chapter 10 quiz
Fourteen questions on Foundry IQ, knowledge bases, data sources and retrieval, and configuring grounded, cited answers. Answer first, then expand Show answer.
Part A · Multiple choice
What core problem does Foundry IQ solve?
- A Slow token generation
- B Building separate RAG infrastructure for every agent
- C Model fine-tuning
- D Image generation
Show answer
B. It provides shared, managed retrieval so multiple agents use one set of knowledge bases.
How does a knowledge base organize information?
- A By storage location
- B By business domain (e.g. Product Documentation, HR Policies)
- C By file size
- D By model
Show answer
B. Domains reflect how people think, and can consolidate multiple sources.
What is the automatic data-source workflow?
- A Discovery → processing → indexing → monitoring
- B Train → deploy → evaluate
- C Map → measure → mitigate
- D Sense → reason → act
Show answer
A. Scan, chunk & embed, make searchable, and reindex on change.
How does an agent connect to a Foundry IQ knowledge base?
- A Through the Model Context Protocol, via an MCPTool
- B By fine-tuning
- C By copying files locally
- D It can't
Show answer
A. The agent retrieves like any other tool, no custom retrieval logic.
Which retrieval behavior is the enterprise standard?
- A Answer from training data
- B Search but don't cite
- C Search + cite + ground
- D Refuse all questions
Show answer
C. Grounded in the knowledge base with source citations.
Which three behaviors should agent instructions specify for reliable retrieval?
- A When to retrieve, how to cite, and fallback behavior
- B Temperature, top-p, and max tokens
- C Model, endpoint, and key
- D Name, description, and parameters
Show answer
A. Always search, cite in an exact format, and handle "not found" gracefully.
Part B · True or false
One Foundry IQ knowledge base can serve multiple agents.
Show answer
True. Improvements to it benefit all connected agents immediately.
Foundry IQ requires you to write custom retrieval code for each agent.
Show answer
False. Retrieval intelligence (query analysis, ranking, citations) is applied automatically.
A well-indexed knowledge base guarantees the agent will cite its sources.
Show answer
False. Without precise instructions, an agent may answer from training data or search without citing.
You should monitor citation and fallback frequency in production.
Show answer
True. Real users ask differently than test scenarios; iterate on the patterns you see.
Foundry IQ is built on Azure AI Search.
Show answer
True. It's a managed knowledge platform layered on Azure AI Search.
Part C · Short answer
Name the four steps of the automatic data-source workflow.
Show answer
Discovery, processing (chunk & embed), indexing, and monitoring (reindex on change).
Name the four built-in retrieval-intelligence capabilities.
Show answer
Query analysis, strategy selection, result ranking, and citations.
Give an example of a source a single knowledge base might consolidate.
Show answer
Any of: SharePoint specs, Azure Blob API docs, OneLake analytics, or existing search indexes.