Foundry IQ: The Knowledge Layer for Agents - Microsoft Ignite 2025
Agents need context. But how should we connect data to our agents for optimal context? Microsoft Ignite 2025 session BRK196 unveiled Foundry IQ—the knowledge layer for agents—and the latest developments from Azure AI Search and Microsoft Foundry. Discover multi-source RAG orchestration, retrieval steering, dynamic security controls, and agentic RAG that delivers 36% higher answer accuracy.
The Context Challenge: Why Agents Struggle Without Knowledge
AI agents without access to enterprise knowledge are like new employees without onboarding—they can perform generic tasks but lack the context needed for meaningful work. The fundamental challenge: how do you connect complex, distributed enterprise data to agents in a way that's secure, performant, and actually useful?
Why Traditional Approaches Fail
Foundry IQ solves these challenges by providing a unified knowledge layer that empowers developers to create intelligent agents capable of accessing, organizing, and interpreting complex enterprise data seamlessly.
Evolution of AI Interaction: From Crafted Queries to Natural Language
The session highlighted a fundamental shift in how users interact with AI:
Traditional Search (2010s)
Users carefully crafted keyword queries: "Q4 2024 revenue report EMEA"
Required: Understanding of document naming conventions, folder structures, and exact terminology
Semantic Search (Early 2020s)
Users asked questions: "Show me fourth quarter revenue for Europe"
Better: Understands intent and synonyms, but struggles with complex multi-step reasoning
Agentic Retrieval (2024-2025)
Users have conversations: "How did we perform in Europe last quarter compared to Asia, and what explains the difference?"
Foundry IQ handles this: Retrieves multiple documents, synthesizes insights across sources, reasons about causation
Foundry IQ is designed to handle this evolution—delivering accurate results even from ambiguous, fragmented, or conversational inputs that would fail with traditional search approaches.
✅ Technspire Perspective: Natural Language Transformation
A Swedish pharmaceutical company's research team spent 40% of their time searching for information across clinical trial data, regulatory documents, and research papers. Their traditional search system required precise queries—researchers needed to know exact document titles, clinical trial IDs, and compound names. When they asked "What safety concerns have we seen with compounds similar to XYZ-123 in cardiovascular patients?", the system returned zero results because it couldn't understand the conceptual relationship or synthesize across multiple trials. After implementing Foundry IQ with agentic retrieval, the same query returned: 7 related clinical trials, 3 safety reports, 12 research papers, and an AI-synthesized summary highlighting the specific cardiovascular concerns found across all sources. Search time dropped from 2.5 hours to 4 minutes per query, and researchers reported finding information they "didn't even know existed in our systems." The natural language understanding transformed information discovery from archaeological excavation to intelligent conversation.
Knowledge Bases: Organizing Multi-Source Enterprise Data
Central to Foundry IQ is its knowledge base system, which organizes information from multiple sources—SharePoint, OneLake, Blob Storage, and publicly available data—without requiring developers to manually manage each source.
Foundry IQ Knowledge Base Architecture
📚 Unified Knowledge Layer
Single abstraction layer over all enterprise data sources—agents query one knowledge base, not dozens of individual systems
🔌 Pre-Built Connectors
Native support for Microsoft 365 (SharePoint, OneDrive, Teams), OneLake, Azure Blob Storage, SQL databases, and 100+ external sources
🔄 Automatic Indexing
Continuous indexing of new and updated content with semantic understanding—no manual schema definition or keyword tagging
🎯 Intelligent Organization
AI-powered categorization, relationship mapping, and metadata enrichment that understands content context and connections
⚡ Real-Time + Indexed
Hybrid approach: fast indexed search for most queries, real-time retrieval when freshness is critical
Azure AI Search Integration: Speed and Depth
Foundry IQ integrates directly with Azure AI Search, supporting both indexed and real-time data retrieval. This offers developers flexibility to balance speed and depth—enabling optimal performance whether the goal is quick discovery or comprehensive data synthesis.
⚡ Indexed Search (Fast)
Use case: Quick lookups, FAQ answers, document discovery
Performance:
- • Sub-second response times
- • Millions of documents searchable
- • Pre-computed semantic embeddings
- • Optimized for throughput
🔍 Real-Time Retrieval (Deep)
Use case: Latest data, complex synthesis, multi-hop reasoning
Capabilities:
- • Queries live databases directly
- • No indexing delay (seconds-old data)
- • Complex joins and aggregations
- • Higher latency but complete accuracy
Foundry IQ automatically decides which approach to use based on the query—or developers can specify preferences for their specific use cases.
Multi-Source RAG Orchestration: Synthesizing Across Systems
One of Foundry IQ's most powerful capabilities is multi-source RAG (Retrieval-Augmented Generation) orchestration—the ability to retrieve information from multiple systems simultaneously and synthesize coherent answers.
How Multi-Source RAG Works
1. Query Understanding
AI analyzes the user's question to identify which data sources are relevant and what information is needed
2. Parallel Retrieval
Simultaneously queries multiple sources (SharePoint documents, OneLake analytics, SQL databases, external APIs) in parallel
3. Relevance Ranking
Scores and ranks retrieved information based on semantic relevance, recency, authority, and source credibility
4. Context Assembly
Combines the most relevant information from all sources into a coherent context window for the AI model
5. Answer Synthesis
Generates comprehensive answer that synthesizes insights across all retrieved information with source citations
Example: Multi-Source Query
User asks: "What were our top 3 product issues last quarter and how did they impact revenue?"
Foundry IQ orchestrates:
- 1. Support ticket system: Retrieves top reported issues from customer support database
- 2. Product analytics (OneLake): Pulls usage data showing feature adoption and error rates
- 3. Financial data (SQL): Retrieves quarterly revenue by product and customer segment
- 4. Engineering docs (SharePoint): Finds root cause analysis and resolution timelines
Result: Synthesized answer identifying the 3 issues, their frequency, affected customer segments, estimated revenue impact, and resolution status—all with citations to source documents.
Retrieval Steering: Guiding the Search Process
Retrieval steering allows developers and users to guide how Foundry IQ searches for information—controlling which sources to prioritize, what time ranges to consider, and how to balance different ranking factors.
🎯 Source Prioritization
Specify which data sources to search first or exclusively:
- • "Search only approved marketing materials"
- • "Prioritize recent engineering docs over archived"
- • "Include external research, then internal analysis"
📅 Temporal Filtering
Control time ranges and recency preferences:
- • "Only documents from last 6 months"
- • "Include historical context but prioritize recent"
- • "Compare current quarter vs same quarter last year"
⚖️ Ranking Adjustments
Fine-tune how results are ranked:
- • Boost authoritative sources (executive communications)
- • Prefer documents with high engagement/views
- • Deprioritize draft or unverified content
🔍 Depth Control
Balance between fast discovery and comprehensive search:
- • "Quick search" mode: Top 5 results, sub-second
- • "Comprehensive" mode: Deep search, 2-5 seconds
- • "Exhaustive" mode: All relevant docs, 10+ seconds
Dynamic Security Controls: Permission-Aware Retrieval
One of the most critical features of Foundry IQ is dynamic security controls—built-in access control and governance that ensures agents only retrieve data users are authorized to see.
⚠️ The Security Risk Without Dynamic Controls
Traditional RAG systems often use a single service account to access all data sources—meaning the AI can retrieve documents the user doesn't have permission to see, creating serious data leakage risks.
Example: A sales agent asks "Show me all customer contracts." Without dynamic security, the AI retrieves contracts from all regions and customers—including ones the salesperson isn't authorized to access—and includes that sensitive information in the response.
✅ How Foundry IQ Enforces Security
✅ Technspire Perspective: Security Prevents Data Leakage
A Swedish financial services company deployed an AI agent for account managers to answer questions about client portfolios. Initially, they used a single service account with read access to all client data—the easiest implementation path. During testing, an account manager asked "What's the average portfolio value of our high-net-worth clients?" The agent returned accurate statistics—but in its explanation, it cited specific client names, account balances, and investment details from clients assigned to other account managers. This violated their regulatory requirement that account managers only access data for their assigned clients. When they implemented Foundry IQ with user-scoped retrieval, the same query returned statistics calculated only from the account manager's assigned clients—never exposing data from other portfolios. The financial regulator audit that would have resulted in a €250,000 fine became a zero-findings pass. Dynamic security isn't a feature—it's the difference between compliant AI and regulatory disaster.
Agentic RAG: Intelligent Multi-Step Retrieval
Agentic RAG represents the next evolution of retrieval systems—where the AI doesn't just search once, but orchestrates multiple retrieval steps, reasons about what additional information is needed, and iteratively builds a comprehensive understanding.
Agentic RAG Workflow
Step 1: Initial Retrieval
Agent performs initial search based on user's question, retrieves first set of relevant documents
Step 2: Analysis & Gap Identification
Agent analyzes retrieved information and identifies gaps—what additional context or data is needed to fully answer the question
Step 3: Follow-Up Retrieval
Agent autonomously performs additional searches to fill identified gaps, potentially from different sources or using different queries
Step 4: Reasoning & Synthesis
Agent reasons about all retrieved information, identifies contradictions, cross-references facts, and synthesizes coherent understanding
Step 5: Answer Generation
Agent generates comprehensive answer that addresses the original question with supporting evidence and citations
Step 6: Confidence Assessment (Optional)
Agent evaluates confidence in the answer and may perform additional verification or flag uncertainty if information is incomplete
Performance Results: 36% Accuracy Improvement
Evaluation metrics demonstrate that Foundry IQ's agentic retrieval delivers significantly higher answer accuracy—up to 36% improvement—over traditional search methods. The system adapts dynamically, optimizing quality against latency for efficient and precise outcomes.
Accuracy Improvement
Over traditional keyword and semantic search methods
Faster Than Manual
Compared to humans searching multiple systems manually
User Satisfaction
In enterprise pilot programs with complex queries
Performance Optimization
Foundry IQ dynamically optimizes the tradeoff between quality and latency:
- → Simple queries: Single-step retrieval, sub-second responses, high accuracy
- → Complex queries: Multi-step agentic retrieval, 2-5 second responses, 36% higher accuracy
- → Critical queries: Exhaustive search with verification, 10+ seconds, maximum confidence
Implementation: Getting Started with Foundry IQ
Organizations can implement Foundry IQ through a structured approach:
Phase 1: Data Source Mapping (Week 1)
Identify key data sources, document permissions, assess data quality and completeness
Phase 2: Knowledge Base Setup (Weeks 2-3)
Connect data sources to Foundry IQ, configure indexing, set up security controls and Purview integration
Phase 3: Pilot Agent Deployment (Weeks 4-6)
Build pilot agent with Foundry IQ integration, test retrieval quality, measure accuracy vs. traditional search
Phase 4: Optimization (Weeks 7-8)
Fine-tune retrieval steering, adjust ranking factors, optimize performance vs. quality tradeoffs
Phase 5: Scale to Production (Weeks 9-12)
Deploy to broader user base, monitor performance and security, expand to additional use cases
✅ Technspire Perspective: Rapid Implementation Success
A Swedish logistics company wanted to build a "Knowledge Agent" for their customer service team to answer questions about shipping policies, customs regulations, and service capabilities across 45 countries. Their previous attempt using traditional search took 4 months and failed—the system couldn't synthesize information from their policy wiki, customs databases, and service manuals distributed across SharePoint, SQL, and third-party APIs. With Foundry IQ, we completed Phase 1-3 in 5 weeks instead of 4 months. The knowledge base automatically indexed 12,000 documents and connected to 3 databases. When a customer service rep asked "What are the import restrictions for lithium batteries shipping to Norway, and what's our fastest service option?", the agent retrieved relevant customs regulations (external API), company shipping policies (SharePoint), and service level data (SQL database), synthesized the answer with citations in 3.2 seconds. Customer service resolution time dropped from 8.5 minutes to 2.1 minutes, first-contact resolution rate jumped from 67% to 91%, and the team reported they could finally trust the AI's answers because it showed its sources.
The Future of Enterprise Knowledge
Foundry IQ represents a fundamental shift in how organizations connect AI agents to enterprise knowledge:
- ✓ Unified knowledge layer: One abstraction over all enterprise data—no manual connector development
- ✓ Natural language understanding: Handle ambiguous, conversational queries that fail with traditional search
- ✓ Multi-source orchestration: Synthesize insights across disparate systems automatically
- ✓ Dynamic security: Permission-aware retrieval that prevents data leakage
- ✓ Agentic intelligence: Multi-step retrieval with reasoning and gap-filling
- ✓ Proven performance: 36% accuracy improvement over traditional methods
As the session concluded, the message was clear: agents without proper knowledge access are severely limited. Foundry IQ provides the knowledge layer that transforms AI agents from generic assistants into intelligent, context-aware partners that truly understand your business.
Ready to Give Your Agents Enterprise Knowledge?
Technspire helps Swedish and European organizations implement Microsoft Foundry IQ to connect AI agents to enterprise knowledge. From data source integration to security configuration and performance optimization, we ensure your agents have the context they need while maintaining compliance and security.
Contact us to discuss how Foundry IQ can provide your AI agents with secure, intelligent access to enterprise knowledge across all your data sources.
Key Takeaways from Microsoft Ignite BRK196
- • Foundry IQ is the knowledge layer for agents—unified access to enterprise data across multiple sources
- • Evolution from crafted queries to natural language conversations—Foundry IQ handles ambiguous, fragmented inputs
- • Knowledge bases organize multi-source data (SharePoint, OneLake, Blob Storage) with automatic indexing
- • Azure AI Search integration balances indexed (fast) and real-time (deep) retrieval based on query needs
- • Multi-source RAG orchestration synthesizes information across disparate systems in parallel
- • Retrieval steering allows fine-grained control over source prioritization, temporal filtering, and ranking
- • Dynamic security controls enforce user-scoped permissions with Microsoft Purview integration—preventing data leakage
- • Agentic RAG performs multi-step retrieval with reasoning, gap identification, and iterative refinement
- • Performance evaluation shows 36% accuracy improvement over traditional search methods
- • System dynamically optimizes quality vs. latency tradeoffs based on query complexity and requirements