Skip to content

Foundry IQ: Document Intelligence

What is Foundry IQ?

Foundry IQ is Azure AI Foundry's unified knowledge layer that enables agents to access enterprise documents through intelligent retrieval.

Key Capabilities

Capability Description
Knowledge Bases Automatic indexing and vectorization of documents
Agentic Retrieval AI-driven search with planning, iteration, and reflection
Enterprise Security Built-in Entra ID authentication and Purview integration
Multi-format Support PDFs, Word, PowerPoint, and unstructured text

How Agentic Retrieval Works

Unlike simple vector search (find similar text), agentic retrieval uses AI to:

User: "What's our policy for notifying customers during extended outages?"

┌─────────────────────────────────────────────────────────────┐
│  Step 1: PLAN                                               │
│  Agent decomposes into sub-queries:                         │
│  • "customer notification policy"                           │
│  • "extended outage definition"                             │
│  • "communication requirements during incidents"            │
└─────────────────────────────────────────────────────────────┘
┌─────────────────────────────────────────────────────────────┐
│  Step 2: ITERATE                                            │
│  For each sub-query:                                        │
│  • Search knowledge base                                    │
│  • Evaluate relevance of results                            │
│  • Refine search if needed                                  │
└─────────────────────────────────────────────────────────────┘
┌─────────────────────────────────────────────────────────────┐
│  Step 3: REFLECT                                            │
│  Before responding:                                         │
│  • Do I have enough information?                            │
│  • Are sources consistent?                                  │
│  • Can I cite specific documents?                           │
└─────────────────────────────────────────────────────────────┘
┌─────────────────────────────────────────────────────────────┐
│  Response with Citations                                    │
│  "According to our Customer Service Policies (page 2),      │
│  customers must be notified within 15 minutes of a          │
│  confirmed outage. The Outage Management Policy (page 1)    │
│  defines extended outages as those exceeding 4 hours..."    │
└─────────────────────────────────────────────────────────────┘

Why This Matters for Customers

Problem: Simple RAG Fails on Complex Questions

Basic retrieval-augmented generation (RAG) does one search and uses whatever comes back. This fails when:

  • Questions have multiple parts
  • Information spans multiple documents
  • The obvious search terms don't match the document language

The agent acts like a knowledgeable employee who:

  1. Understands what's really being asked
  2. Knows to check multiple sources
  3. Reconciles conflicting information
  4. Admits when it can't find an answer

Customer Talking Points

Question Response
"Why not just use search?" "Search finds documents. Agentic retrieval finds answers — and knows when to look in multiple places."
"What about hallucination?" "Every response cites specific documents. Users can click through to verify. The agent says 'I don't know' rather than guess."
"Can it handle our complex policies?" "The Plan-Iterate-Reflect approach handles multi-part policies. Let me show you with this example..."

Technical Details

Document Processing Pipeline

PDFs/Word/PPT → Text Extraction → Chunking → Embedding → Vector Index
  • Chunking: Preserves sentence boundaries, typically 500-1000 tokens
  • Embedding: Azure OpenAI text-embedding-3-large (3072 dimensions)
  • Index: Azure AI Search with hybrid (keyword + vector) retrieval

Search Configuration

# Hybrid search combines:
# 1. Vector similarity (semantic meaning)
# 2. Keyword matching (exact terms)
# 3. Semantic ranking (re-ranking for relevance)

query_type = "vectorSemanticHybrid"

← Overview | Fabric IQ: Data →