Slides: GraphRAG: The Marriage of Knowledge Graphs and RAG: Emil Eifrem

Source Video

GraphRAG: The Marriage of Knowledge Graphs and RAG: Emil Eifrem

Relationship To World's Fair 2026

These slides are extracted from a public AI Engineer YouTube video connected to World's Fair 2026. Speaker-matched clips are supporting context unless later confirmed as exact session recordings; official livestream recordings are day-level/event-level source material.

Related Scheduled Sessions

Extracted Slides

slide-001.jpg

OCR text:

INNOVATIONSPONSOR

aws

PLATINUMSPONSORS

MongoDB.

Google Cloud

neo4j

slide-002.jpg

OCR text:

of

we

- .

e04

Rave bl oo

a

slide-003.jpg

OCR text:

The Evolution of... Web Search

Serious Sports Fans On)y_$1 000.000 in Cash and Pozes!

For serious sports fans only! Play Fantasy Football!

Sa

si "it's amazing where

Go Get It will get you.

Find: [ Go Get tt j

Enhance your search.

New Scarch » TopNews « Sites by Subject «Top 5% Sites - City Guide - Pictures & Sounds

PeopleFind - Point Review - Road Maps - Software - About Lycos - Club Lycos » Help

Add Yout Site to Lycos

Copynght € 1996 Lycos™, Inc. All Rights Reserved.

: Lycos is a trademark of Camegic Mellon University.

slide-004.jpg

OCR text:

The Evolution of... Web Search

. Full text Era: 1994 - 2000

|

nae 4s Netscape

| TT cela) re lve s 4 AOL

7 e we” Rie

- ma llelcokxe aS STOO UMMC LS

slide-005.jpg

OCR text:

a

Po

7

Ta civ)

a | World's Fair

2

slide-006.jpg

OCR text:

Google Launches World’s Largest Search Engine & Subscnde

: Google Now Enables internet Users to Search More Than 1 Billion URLs, Providing Quick and Easy Access to 560 Million Full-Text

Indexed Web Pages and 500 Million Partlaity indexed URLs

MOUNTAIN VIEW, Calif. - June 26, 2000 - Google Inc., one of the fastest growing search engines on the web, today announced it has

released the largest search engine on the Intemet. Google's new index, comprising more than 1 dilbon URLs. offers users the wed's most

comprehensive collection of websites, which can be easily searched with Google's fast and highly relevant search lechnology. Available now

at www.goog'e com, Google's portal and destination site customers can also bcense this new index for integration with their own websites.

we Google is based on a variety of innovative

as technologies, including sophisticated text matching

and its advanced, patent-pending technology called

PageRankim, which ensures that the most important ;

a results always come up first. ;

PT os : .

/ WT at ° aws

; a Microso

Pg

slide-007.jpg

OCR text:

The Evolution of... WWeb Search

. ; egelet-1ne-] 0a off: eae) 4 Oa

nas Bel Niger.) 01)

Search ve Rrra vy I

7 ° ©

: a Microsoft Ea? awWws

slide-008.jpg

OCR text:

t

d

se _

E _ 7

SS

x | Worta's Fair

a

; co oo

. a"

slide-009.jpg

OCR text:

Geogle Se mnere cow saga & € 3 |

Mow cre Comme — ) ™ .

a newer a oo ® eo

Moscone Centar Homepage So -

FRAO ee RR RE Cm apr cirneg ktenetie cote

SELENE RE Seng LE Nagra cen Mmaton lecterens&

Everts

Cet cane ete ee Le ge Ca heen Mare te

Moscone Center

Floor Plans a ie pelese

aie ee a bee Bese Man ee teeth eae bee ember ohne

Explore the Neighborhood Kirtan QS ivetire Tunes

Neat Mylene beg bri tien cated e he

Gerechons and Parking oe

BAST me tee ce reeerd ong tom teres eet AND

Contact Us ta Serpe Mere arian ere py en re

ee EE ARATE HOSES Mow ite ne Ma le ead teem aes eM evPaberaad congas e Ne

Viste etleme ted heey OM oregano id ee tee

Rare matte hor mcacane come © ate tet at mse vee fee te wd ees Fa ast od Monat

reves Wepeee

p also ask Mamete 10 eed Ste bene CAMEL

Penne ty petra neue

wescek Man Tiina: aseeBe Corus? . Cree Cy rd manta Nan Prem

Peer <8 $198 ace

Wo owes Ihe Micecone Carter? 7 Fea Ce Dee

Conmenenmen weed $18) names Megane Sie TE mace,

~ oe

Pea}

WAL a SiMOL eS

slide-010.jpg

OCR text:

- eer gz ve

Moscone Cente Fy sal

bee eT Bs

~ Sor eee 7 Moscone —

“ (Oferaitrg

cass reece so an er

Panag Po merry arm i+ ees rv a

moe suena E . ; a ;

ee an ra ae

[6Telol asic) at

Vener Moscone

v: ae

ry <a

- j © aws

: a Microsoft SOO

slide-011.jpg

OCR text:

Google1/o'24Keynote

AIE

gqpn

sqypeenun

freshsefoodCanyoupullmy figtand htelifo

LG

alland heipme plan the wek

erd?

erved2024

Microsoft

smol

aws

slide-012.jpg

OCR text:

What IS GraphRAG?

ct avacy lanl elon

Clea ]e)nl ne: (Ci cee aun era mn ere a TeTEREOEE SS

a Microsoft Ea? awWws

slide-013.jpg

OCR text:

User

Your Application |

slide-014.jpg

OCR text:

GraphRAG Retrieval Patterns

1. Doa vector search to find an initial set of nodes

slide-015.jpg

OCR text:

GraphRAG Retrieval Patterns

Oates ot0 ac Ras bette ee eee SNe ee ee eae te cree ee ae

|. Do avector search to find an initial set of nodes

2. Traverse the graph around those nodes to add context

3. (Optional: Rank the results using the graph and pass the

top-k documents to the LLM)

| ; )

Ma Iracesxe) aS OVO UMEEGLU

slide-016.jpg

OCR text:

S The Benefits of os

GraphRAG oe

: a Microsoft GiQU° aws

slide-017.jpg

OCR text:

‘O) Higher ane Tin)

X the accuracy of LLM responses

by 54.2%, an average of 3x.

Accuracy

The clearest driver of a aan Sequeda

GraphRAG adoption amongst s

users Is higher accuracy. ows lunes RE REEL AO

wt A

0 | a (

Ce ee

a "1 Sy" my

EGG Deca sueae denne Oe ao

GR AsIdeL AAGOUEIG. HE Sin eral EQL DEVON Laas, ike a She

ae _ Tee yt ep,

Ss] 2=.. ==. ee

apg

; 14 me Oats

= , wee

Microsoft Sic? aws

slide-018.jpg

OCR text:

ital

Accuracy

GraphRAG: Unlocking LLM

Ree eld discovery on narrative private data

GraphRAG adoption amongst Renres dire ‘Si

users is higher accuracy.

“By combining LLM-generated

vate ert knowledge graphs and graph

Oe w . ,

SE EA machine learning, GraphRAG

Beet Fea enables us to answer important

Pett Ph kept 5

a er classes of questions that we cannot

ek .

une attempt with baseline RAG alone."

: - BE Microsoft

slide-019.jpg

OCR text:

Devel t |

The second reason we hear

people choose GraphRAG over

vector-only RAG is easier

development... once they've

pushed through the initial

learning curve.

" Werdly ease of develapment Jer lack

thereof! :s aiso ene of the stumbins: blocks!

How can that be?

One word Knowledge Graph Construction’

Ok that’s three. Susneld that tnouenAtuerd

ae

slide-020.jpg

OCR text:

@ Easier Development: Why?

Natural Language Description: “Apples and oranges are both fruits”

DATA SEMANTICS

@ Representation

Pe ry ai)

MA Iracesxe) | aSONDOU MEELIS

slide-021.jpg

OCR text:

oe eo ee os BOG act ee

{ 3.91906$9e-03) 2 665977 %e 0] 1 OPI -O3 2 TEE od

399778 Oe Cd 3 O44 ee} 1 DSR 1 450690803

6.4 96e0-04 FOIE 1Ge 0} 2.197 7584e03 $ 14801 726-05

3 NAPI270e-03 19477583402 6 B27eMH7e D4 7 713 7ChEe OF

2 SNES? 7 Ah ete Ob 6 OFS 2c 4 4M Se03

+16 7OEOBBe 3 1 -74174256-0) 2 421641360} 365457 e-0)

1 9O7123%0 03 2S4GM ILC OF} | TAIZ 0) 49) 74S OE

. 3974s PO4e-03 2 21 Me OF) 0541 9e O41 6.724542 0)

~ . # 2.LOP*SSbe 03 2 1624754 OT 2 1620055 OH 1.600051 03

. $ A ORAM ee 0) 4 ISSO OF 3 SO7KS MeO) 4 A2E He 0}

8 2.87379) be 03 45569 Whe OF 7687 OE IL! 0]

TSP Ne Od 4 242d rede 1) FRIIS I) 4140 ee ON

LOOM le OF 5 2IMSO3e O46 $ BIS e OF 4 MIG SEO)

+ zo 2497208 OC} 4 BOO MeO) 2 W141 4e 0) BORA 1 SeOe

¥ Z1LGS2S he CS 8.162] Pode Ob 281272916 03 6.748 2e OF

sa ® ® LIPLWO7Pe-O3 J OLS9 4B! G3 1 71 Bae 03-1 WII)

‘ t 2.96374 360-03 3.37 P6420] 2 2b ea20e OF | BAS) eee OF

+2 AT1LBS2ILE-O) 1 9754141003 2 G10M Me G3 2.2 sesh Od

2AM] TC? L.2O1FS2e OF 3 9961869 03 4 04191 02¢ 03

2 CSORB2 eG) 4 S89 BBe-O} 49599132641) +) OFS 22e-63

258672630} 3 W61310r 03 2 FM OI 4 87257 CT

6 MAD eG 3 Fate Tle O) 4 28959 let 6 TIKI LeOF

12614765 03 & SOLA ihe OF 2 Soe OF 10237564 OD)

4502 4ob De 0} 4 TSO EO) -4 26591008 0) 4 O81 O1Le-0)

392145520 OF 2 42621200 03 8 11921 6he OS 41112076" 0)

ie ee Ce aaa ee ee ae ee eee ee

Band

” Microsoft Smoln aWws

slide-022.jpg

OCR text:

,

@ Easier Development: Why?

:

eet een ee eee ee ct ats 7

[ 3.91508$90-03 2.6659777e-03 1.0298982e-03 -2.7156321¢-03

1.99778708-03 3.120442462-03 1.205S682¢-04 1.0450699¢-03

6.4308 2960-04 3.0822$198-03 2.19725540-03 $.1480172e-05

+3.7099270e-03 3.9439583¢-03 6 8276987 04 7.713 70660-04

2.36985200 -03 -7.8547641¢-04 6.0383842¢-04 4.6370425¢-03

-1.6786088e-03 1.7417425¢-03 2.4216413¢-03 3.6545738e-03

. +1.98712390-03 2,9489421¢-03 -1.2810022¢-03 -4.91740532-04

‘ . ’ ‘ -3.9743204e-03 -2.7023794e-03 -3.05419500-O4 -1.5724347e-03

z » ‘ , 2 -2.1029566¢-03 -2.1624754e-03 2. 162005Se-04 -1.4000515¢-03

3 , ~4.082486Se-03 4.6588355e-04 3.5028579e-03 4.8283348e-03

: -2.8737928e-03 -4.5569206e-03 -7.6568732e-04 -3.3311991e-03

E s 3 sort . 3.$790715e-03 4.2424244¢-03 3.347822Se 03 -7.41403960-04

1.00301 1L1e-03 -$.2394503404 5.8383477e-04 -4.643099S¢-03

Lo 7 2.6972082€-03 -4.8002079e-03 -2.3013414e-03 §.0388715Se04

. . 3.395257Se-05 -8.1621204¢-04 3.812729 1€ 03 6.742820 04

a : . ~ 1 -1.7713077e-03 -3.0159748e 03 1.7178850e 03 -1.9258332¢-03

: : s od -2.4637436e-03 3.37796S2e-03 2.7676420e-03 1.8853768¢-03

. ‘ . -2.4718521€-03 -2.9754141e-03 2.61040360-03 -2.1335895¢-03

2.4405 3340-03 -3.2013952e 04 3 99618690-03 4.0419102¢:03

, . a 2.0586823¢-03 4.9897884e-03 4.55991 324-03 -1.0976522¢.03

1.$563263e-03 3.9063310e-03 -2.9308300e-03 4.8254002e-03

-8.7642738¢-06 3.974867 1e-03 $ 2895391e-04 63330121606

+1.2614765¢.03 -8.5018738e O4 3.7659388e.03 3.0237S64e 03

4.5014662€-03 4.32587934-03 -4.26591008-03 4.9081761¢-03

-3.9214852€-03 -2.4262110e-03 -8.1192264e-05 -4.11120764-03}

Customer “i actualy already Pxed a couple of bugs thanks te this!”

slide-023.jpg

OCR text:

Knowledge Graph Construction

Typically PDFs or other phe ee Structured data with

text documents Structured data with short text values

long-form text

Bee -_ as

| a

an a.

. . if fai)

a Microsoft @aU? awWws

slide-024.jpg

OCR text:

Two Types of Knowledge Graphs

A graph representation of for

example the words, paragraphs.

chunks, documents and the

relationships between them

1 . Pa

; = mE lrelxekxe) aS O00 O UMMC LS

Slide-Derived Subjects To Review

Subject extraction uses video title, related session titles/descriptions, transcript context, and OCR text when available. OCR is best-effort and should be reviewed against the embedded slide images.