Skip to main content
Microsoft IQ
Foundry IQ · Web IQ · Work IQ · Fabric IQ
The context engine for enterprise agents
Udaiappa Ramachandran (Udai)
CTO|CSO @ Akumina Inc. | Microsoft MVP (AI)
Web: https://udai.io
LinkedIn: https://linkedin.com/in/udair
Meetup: https://meetup.com/nashuaug
What this deck covers
One platform, four intelligences — read at three altitudes.
Foundry IQ Agentic retrieval over your knowledge + live web
Web IQ An agent-native index of the connected web
Work IQ Grounding + action across Microsoft 365
Fabric IQ A shared model of how the business operates
Three reading lenses
Executive
Why it matters: fewer silos, grounded answers,
governed automation.
Architect
How it fits: protocols (MCP · A2A · REST), data flow,
and boundaries.
Developer
How to build: SDK calls, tools, and deployable agent
patterns.
2
Microsoft IQ
The Microsoft IQ platform
A context layer that sits between your agents and everything they need to reason over.
Copilot & custom agents
context in · actions out
Foundry IQ
Agents unlock knowledge
Policy-aware retrieval over docs + web
Web IQ
Connect to the web
Fresh, verifiable web intelligence
Work IQ
How employees work
People, mail, meetings, chats, files
Fabric IQ
How the business operates
Entities, metrics, systems of record
Your enterprise: Microsoft 365 · OneLake & Fabric · documents · the open web
3
Microsoft IQ
S E C T I O N 0 1
FOUNDRY IQ
Agentic retrieval that turns your knowledge — and the live web — into grounded, cited answers.
FOUNDRY IQ WEB IQ WORK IQ FABRIC IQ
4
Microsoft IQ - Foundry IQ
Making knowledge agent-ready
Knowledge lives everywhere. Foundry IQ turns it into one grounded answer.
PDFs Email
Teams
chats
Tables
Analytics SharePoint
Foundry IQ
built on Azure AI Search
1. Route
2. Retrieve
3. Merge
4. Synthesize
One grounded answer
Semi-gloss is best for bathrooms —
moisture-resistant and washable [1]. A can
of ZavaTech Interior Semi-Gloss runs $47.00
[2].
One question. Every source. One answer — with citations.
5
Microsoft IQ - Foundry IQ
Inside a knowledge base
Plan queries, search every source in parallel, then merge and synthesize — iterating only when it still needs more context.
Input Knowledge sources Output
Conversation
Steering instructions Retrieval planning
Search query 1
Search query 2
Search Index
MCP
Work IQ
Fabric IQ
Merged results
Activity log
Answer synthesis
Iterative retrieval
Microsoft IQ - Foundry IQ 6
Getting knowledge in
Unstructured files run an ingestion pipeline. Choose how much control you need.
Extraction
text, tables & figures
Chunking
split by sentences
Embedding
vector per chunk
Index
one row per chunk
File Knowledge Source
You bring files — zero ingestion code
Extract → Chunk → Embed → Index, fully automatic
Just upload and go
No pipeline to maintain
Fixed index schema, less flexibility
Best for: prototyping, small document sets
Indexed Knowledge Source
You bring a data source (Blob, DB, OneLake…)
Data source → Indexer → Skillset → Index
Full schema & pipeline control
Custom enrichments / skillsets
More custom code to maintain
Best for: production data, custom schemas
7
Microsoft IQ - Foundry IQ
Quality knobs: hybrid search + effort
Fuse keyword with vector, then dial how hard the engine reasons per run.
Hybrid retrieval
Hybrid beats pure vector or keyword alone.
Keyword results
exact matches, BM25
Vector results
semantic similarity
Fusion (RRF)
Semantic re-ranker
Retrieval reasoning effort
One knob trading cost & latency for depth.
MINIMAL
No LLM — re-ranker only. Lowest cost & latency; you plan queries yourself. Ideal as an
agent tool.
LOW
LLM plans queries + optional answer synthesis. Balanced default.
MEDIUM
Adds a second retrieval pass only when more context is needed. Highest quality.
more agentic, more cost →
8
Microsoft IQ - Foundry IQ
Using Foundry IQ inside agents
Three ways to wire a knowledge base into an agent — pick by control vs. portability.
Option 1
Custom tool via API
Full control over params & response
Access to references + activity log
Less portable across agents
Option 2
Point at KB MCP endpoint
Least code to maintain
Portable, standards-based
Can't tune retrieval params
Option 3
Via Foundry Toolbox
Easy to add more tools
Managed auth for authorized sources
Inherits MCP limitations
@tool
async def retrieve_company_knowledge(queries: list[str]):
request = KnowledgeBaseRetrievalRequest(
intents=[SemanticIntent(search=q) for q in queries])
return await knowledge_base_client.retrieve(request) # grounded + cited
Repos & notebooks: github.com/microsoft/iqdeepdive
9
Microsoft IQ - Foundry IQ
S E C T I O N 0 2
WEB IQ
An agent-native index of the open web: fast, structured, multimodal — delivered in a single call.
FOUNDRY IQ WEB IQ WORK IQ FABRIC IQ
10
Microsoft IQ - Web IQ
An agent-native view of the web
Not a search box for people — a structured, multimodal web index built for agents.
Web IQ
Web search
News
Finance
Sports
Video
Browse
…with more verticals to come. One index, many modalities.
11
Microsoft IQ - Web IQ
Why an agent-specific web engine
Built for the way agents consume the web: fast, and everything in one payload.
~164ms
P95 LATENCY
fast enough to sit inside a reasoning loop
2.5×
FASTER
than today's best alternative
1 call
FULL PAYLOAD
Markdown, HTML & summaries together
Latency is a feature An agent may call the web many times in one turn. Round-trip cost compounds — so Web IQ trims it aggressively.
One response, no second hops Human search returns links you must open. Web IQ returns the content — text, structure, and a summary — in the first response.
Fresh & verifiable Results carry URLs, crawl timestamps and language, so agents can ground answers in citable, current sources.
12
Microsoft IQ - Web IQ
Retrieval through the Web IQ MCP server
The agent calls one `web` tool; Web IQ returns a structured, citable result.
Agent question
“Industry benchmarks for mental-health
benefits at tech companies?”
Web IQ
MCP server
Structured result
Tool web
Title Mental Health Technology Sector Overview
URL multiples.vc/coverage/mental-health-technology
Content Telehealth platforms, measurement-based care, practice-
management software…
Crawled 2026-05-10T20:07:00Z
Language en
Private preview
Allow-listed customers only — a WEB_IQ_KEY is required to call the MCP
server.
13
Microsoft IQ - Web IQ
Web IQ inside a Foundry knowledge base
Register Web IQ as an MCP source, then blend live web with your own documents.
Web IQ
live web results
HR docs index
your documents
Knowledge
base
One blended answer — company policy + current web, ranked together and
cited.
web_source = McpServerKnowledgeSource(
name="web-knowledge-source", description="Web IQ (live web search)",
mcp_server_parameters=McpServerKnowledgeSourceParameters(
server_url="https://api.microsoft.ai/v3/mcp", tools=[McpServerTool("web")]))
kb = KnowledgeBase(knowledge_sources=[web_source, hrdocs_source]) # blend both
14
Microsoft IQ - Web IQ
When to reach for Web IQ
Architect's checklist: it grounds on the open web, not your private estate.
Reach for it when
The answer depends on current, public information
You need news, finance, sports, or web media
You want citations to verifiable external sources
Latency inside a reasoning loop matters
Use something else when
The answer lives in your documents → Foundry IQ
It's about people, mail or meetings → Work IQ
It's business / operational data → Fabric IQ
Access is gated → still in private preview
Mental model: Foundry IQ is the retrieval brain; Web IQ is one of its most powerful sources — the world's knowledge, made agent-shaped.
Because it plugs in as an MCP knowledge source, the same knowledge base can answer from your files and the live web in a single call.
15
Microsoft IQ - Web IQ
S E C T I O N 0 3
WORK IQ
A routing orchestrator over Microsoft 365 that lets agents retrieve context and take action in
place.
FOUNDRY IQ WEB IQ WORK IQ FABRIC IQ
16
Microsoft IQ - Work IQ
An agent for your work data
A routing orchestrator over Microsoft 365 — retrieve context and take action, in place.
Your agent
or app
MCP
A2A
REST
Work IQ routing
orchestrator
Context API Tool APIs
Microsoft 365
Mail Calendar Teams
Docs People SharePoint
“What did my manager email me this week?” → Niya sent 3 emails on the Q3 budget: she needs your headcount by Friday and moved the sync to Thursday
2pm. [1][2][3]
17
Microsoft IQ - Work IQ
Work IQ vs. Microsoft Graph
Graph is a data-access API. Work IQ is an intelligence & context layer for agents.
Microsoft Graph Work IQ
A data-access API An intelligence & context layer
Built for apps and users Built for agents
App-only & delegated auth Delegated (signed-in user) only
Returns raw data you stitch together Returns grounded answers + references
Hundreds of typed endpoints 10 generic tools over MCP
You manage paging & throttling Agent-optimized, in-place governance
18
Microsoft IQ - Work IQ
One layer, three protocols
Same intelligence, exposed through the protocol that fits the caller.
Chat
Converse with people & other agents —
ask, A2A
Context
Grounded org understanding across mail,
meetings, docs & chats
Tools
Governed retrieve + act across M365 via
generic verbs
Workspaces
A persistent place for agents to hold state
across long tasks
A2A Agent → Agent
One agent delegates a task to another (multi-
agent orchestration)
MCP Agent → Tool
An agent grounds & acts via Work IQ as a tool
REST Human / device → Agent
A mobile or web app queries Work IQ for a user
19
Microsoft IQ - Work IQ
10 generic tools — and one write path
Progressive disclosure: discover, learn the schema, then read or (carefully) write.
Progressive disclosure
1
Chat
ask · list_agents
2
Schema
search_paths · get_schema
3
Entity read
fetch · call_function
4
Entity write
create · update · delete · do_action
do_action
the ONLY write path
ALLOWED PREFIXES
/me/ /users/ /sites/
BLOCKED
/authentication/ /servicePrincipals/
Discover first: search_paths → get_schema
Runs under the signed-in user
jsonBody is a JSON-encoded string
20
Microsoft IQ - Work IQ
From tool to autonomous coworker
Wire Work IQ as an MCP tool — then ship the same agent as an Agent 365 autopilot.
work_iq = MCPStreamableHTTPTool(
name="work-iq", url=work_iq_endpoint,
allowed_tools=["ask", "fetch",
"get_schema", "do_action"])
agent = Agent(name="ContosoWorkmate",
tools=[get_current_date, work_iq])
Ship it as an Agent 365 autopilot
Its own M365 identity — mailbox, calendar, Teams presence
@mention it in Teams like a coworker
Reasons over real work context via Work IQ
Governed like any identity — auditable & revocable
In the playground → runs on-behalf-of YOU
Uses your mailbox, your permissions, your identity.
In Teams → acts as its OWN identity
do_action takes real action from the autopilot's M365 — think “digital coworker,”
not “assistant signed in as me.”
21
Microsoft IQ - Work IQ
S E C T I O N 0 4
FABRIC IQ
A shared business model over OneLake — ontology, graph, and semantic data agents in one layer.
FOUNDRY IQ WEB IQ WORK IQ FABRIC IQ
22
Microsoft IQ - Fabric IQ
Unify data in OneLake with zero ETL
Shortcut & mirroring sources — connect data where it lives, no copies.
GENERALLY AVAILABLE
Azure SQL MI Azure Blob Storage Azure Cosmos DB
Azure PostgreSQL SQL Server 2025 SQL Server
Azure SQL DB Azure Data Lake Store Microsoft OneLake
Snowflake Google Cloud Storage Microsoft Dataverse
Databricks Catalog Amazon S3 S3 Compatible
PUBLIC PREVIEW
SharePoint /
OneDrive
SAP Datasphere
Oracle DB Google BigQuery
Live web of enterprise sources — more added each
release.
Microsoft IQ - Fabric IQ 23
A shared model of the business
Over one data foundation (OneLake), Fabric IQ layers the meaning agents reason with.
OneLake — one unified store for structured, unstructured, real-time & graph data (zero-ETL shortcuts & mirroring)
Ontology
Shared vocabulary of entities &
relationships
Graph
Connections between entities &
events
Semantic model
Curated metrics & KPIs for Power BI
Data agent
AI analyst answering domain
questions
Business intelligence · operations · planning · digital twins all sit on the same shared context.
24
Microsoft IQ - Fabric IQ
Ontology: your business vocabulary
Entity types, properties, typed relationships — bound to real data in OneLake.
Entity types
The concepts (Product, Store, Category)
Properties
Named facts (Product.name, Product.sku)
Relationships
Typed links (Category → contains → Product)
Data bindings
Connect to OneLake (Store.storeName →
store_name)
Retail store ontology
Store
storeId
storeName
city
Inventory
storeId · sku
quantityOnHand
Product
name · sku
price
Category
categoryName
Supplier
supplierId
holds recordsStockFor
contains
supplies
create-lakehouse.py
Microsoft IQ - Fabric IQ 25
Graphs: model relationships, query with GQL
Nodes are things, edges are relationships — traversed with graph queries.
Product
name ·
sku
Review
user_id
Feature
name
HAS_REVIEW
MENTIONS
Nodes, edges, labels & properties
Edge properties add business context — e.g. sentiment (+/−), confidence.
MATCH (p:`Product`)
-[:`HAS_REVIEW`]->(r:`Review`)
-[m:`MENTIONS`]->(f:`Feature`)
FILTER m.sentiment = 'negative'
RETURN p.name, r.rating,
f.featureName, m.confidence
ORDER BY m.confidence DESC LIMIT 20
From agents: add the graph to a Fabric data agent for NL2GQL.
26
Microsoft IQ - Fabric IQ
Semantic models & the data agent
Star-schema models power Power BI and NL2DAX; the data agent queries across it all.
Star schema (web analytics)
Sessions
FACT
Date Devices
Channels Pages
Geography
Facts to aggregate · dimensions to group & filter. Measures = reusable DAX.
Fabric data agent
One conversational agent that selects a source, generates the query, and validates it:
Semantic model DAX
Lakehouse SQL
Graph model GQL
Eventhouse KQL
Data warehouse SQL
27
Microsoft IQ - Fabric IQ
Fabric IQ inside a Foundry knowledge base
Plug operational data in as a source — agents answer over docs and business data together.
“What are the stock levels
for our hammers?”
Knowledge base
Fabric IQ source
HR docs index
Grounded, live answer
12″ Hammer (HM13005): 50 on hand at STORE-SEA · 6″ Hammer
(HM16325): 25 on hand
Ontology
Product inventories
Graph
Product reviews
Semantic models
Website analytics
Data agent
All of the above, unified
A Fabric data agent knowledge source can expose ontology, graph & semantic models through one connection.
28
Microsoft IQ - Fabric IQ
How the four compose
Foundry IQ is the retrieval brain; the others are sources it can blend in one call.
Your agent
Foundry IQ
agentic retrieval — plan · merge · synthesize
Web IQ
the open web
Work IQ
M365 work context
Fabric IQ
business data
Add or remove a source without re-architecting the agent — one knowledge base, many kinds of context.
29
Microsoft IQ
References
Docs, samples, hands-on training, and session recordings.
1
Microsoft IQ series
https://github.com/microsoft/iq-series
Overview and docs spanning the four IQs.
2
IQ deep-dive — samples & notebooks
https://github.com/microsoft/iqdeepdive
Runnable code for the demos in this deck.
3
Get started with Foundry IQ
https://learn.microsoft.com/en-us/training/modules/get-started-foundry-iq/
Hands-on Microsoft Learn training module.
4
NH Cloud on YouTube
https://youtube.com/@nhcloud
Recordings of all our meetings and sessions.
Microsoft IQ 30
Q & A
Executives
Microsoft IQ turns four kinds of context into grounded, governed answers — less silo-stitching, more trustworthy automation.
Architects
One retrieval brain (Foundry IQ) blends web, work and business sources over MCP · A2A · REST — governed in place.
Developers
A few SDK calls register a source and a knowledge base; the same agent ships to Teams as a governed autopilot.
Code, notebooks & demos: github.com/microsoft/iqdeepdive · Requires an Azure account (gpt-5.4 + text-embedding-3-large).
Take it back to your team — three lenses, one platform.
31