Beyond Limits: How GraphRAG Revolutionises Data Interaction
The document discusses the evolving landscape of generative AI and the role of retrieval augmented generation (RAG) and knowledge graphs in enhancing AI applications. It highlights the importance of integrating domain-specific knowledge to improve accuracy, explainability, and reliability of AI-generated responses, particularly in enterprise contexts. Additionally, it introduces 'graphrag' as a composite approach that merges graph analytics with generative AI to enrich data retrieval and contextual understanding.
Presentation on GraphRAG concept by Kristof Neys including an outline of key topics.
Discussion on the state of Generative AI highlighting its good, bad, and ugly aspects and challenges faced.
Identifies key challenges with Generative AI including lack of knowledge, verification issues, and bias.
Introduction to RAG, its definition, and how it enhances language models by providing relevant contextual data.Discusses shortcomings of vector databases in terms of data leverage, context capture, and explainability.Describes how Knowledge Graphs provide structured representations of facts and enhance data context.
Details on how combining vectors with knowledge graphs improves data search relevance and capability.
Presentation on GraphRAG as an advanced technique for understanding text datasets through improved methodology.
Discusses the need for improvement in RAG applications and hints at innovative approaches.
Explains the intersection of data science and graph technology to improve efficiency and insights.Example of applying GraphRAG in RFP generation to streamline the process and improve efficiency. Highlights how Neo4j enhances RFP management by addressing challenges and streamlining data handling.
Presents mechanisms of ensuring responses in GenAI applications are accurate, contextual, and explainable.
Details on utilizing graph algorithms in data science to improve the relevancy of document retrieval.
Wrap-up of the presentation covering the capabilities of GraphRAG and thanking the audience.
Outline
1) Brave NewWorld…
2) Why RAG?
3) Can Knowledge Graphs help?
4) From RAG to GraphRAG
5) Improve your R
6) An example….
7) Does it work?
Neo4j Inc. All rights reserved 2023
2
GENAI
GENAI
The State ofGenerative AI
The Good 💪 The Bad 👎 The Ugly 😱
GenAI Alone != Right Outcomes 🤯
GENAI
12.
Challenges with GenAI:Stochastic Parrot?
● Lack of enterprise domain knowledge
● Inability to verify answers
● Hallucination
● Ethical and data bias concerns
● and more
12 Neo4j Inc. All rights reserved 2024
GenAI
PARROT
13.
13
Managing AI risk
isthe biggest
barrier to scaling
AI initiatives1
Skepticism: Over half of business leaders are
skeptical in adopting GenAI.2
Neo4j Inc. All rights reserved 2024
Explainability: Over 80% of executives worry
about non-transparent nature of GenAI could
result in poor or unlawful decisions.2
Reliability: Inaccuracy and hallucination are two of
the most-cited risks of adopting GenAI
technology at all levels of an organisation.3
1. Deloitte’s State of AI in the Enterprise 2. BCG’s Digital Acceleration Index Study 2023 3. McKinsey: The state of AI in 2023
14.
14 Neo4j Inc.All rights reserved 2024
How can enterprises use
domain-specific knowledge
to rapidly build accurate,
contextual, and explainable
GenAI applications?
Problem
Statement
Retrieval Augmented Generation:
Theability to dynamically query a large
text corpus to incorporate relevant factual
knowledge into the responses generated
by the underlying language model
17.
Neo4j Inc. Allrights reserved 2023
17
RAG augments LLMs by retrieving up-to-date,
contextual external data to inform responses:
Retrieve - Find documents of interest for the
user question
Augment: Combine the user question with
the relevant documents
Generate: Feed enhanced prompt to an LLM
and obtain answer
Retrieval Augmented Generation
Database of Truth
RAG is becoming an industry standard
18.
Why RAG WithVector Databases Fall Short
Similarity is insufficient for rich enterprise reasoning
Neo4j Inc. All rights reserved 2024
18
1
3
2
4
Only leverage a fraction of
your data: Beyond simple
“metadata”, vector databases
alone fail to capture relationships
from structured data
Miss critical context: Struggle to
capture connections across
nuanced facts, making it
challenging to answer multi-step,
domain-specific, questions
Vector Similarity ≠ Relevance:
Vector search uses an incomplete
measure of similarity. Relying on it
solely can result in irrelevant and
duplicative results
Lack explainability:
The black-box nature of
vectors lacks transparency
and explainability
Recap a KnowledgeGraph
A knowledge graph is a
structured representation
of facts, consisting of
entities, relationships and
semantic descriptions
20 Neo4j Inc. All rights reserved 2024
MICA
ANDRE
Name: “Andre”
Born: May 29, 1970
Twitter: “@andre123”
Name: “Mica”
Born: Dec 5, 1975
CAR
Brand “Volvo”
Model: “V70”
Description: ‘Blue external, red seats’
#days: 5/7
LOVES
LOVES
LIVES WITH
O
W
N
S
D
R
I
V
E
S
Since:
Jan 10, 2011
Now with Vectors!
Vectorsas Node properties
=
Vector Search + Graph
Traversal
22 Neo4j Inc. All rights reserved 2024
MICA
ANDRE
Name: “Andre”
Born: May 29, 1970
Twitter: “@andre123”
Name: “Mica”
Born: Dec 5, 1975
CAR
Brand “Volvo”
Model: “V70”
Description:
#days: 5/7
LOVES
LOVES
LIVES WITH
O
W
N
S
D
R
I
V
E
S
Since:
Jan 10, 2011
23.
Neo4j Inc. Allrights reserved 2023
23
Neo4j - Vector Database Capabilities
Vector Search Data Science
Knowledge
Graph
● Find nodes using an implicit similarity search in
the vector index* and enrich with additional
explicit relationships from the knowledge graph
● Hybrid Search with text
● Create vectors of network information using
node embeddings
Now a top 10 vector database on LangChain.
24.
Neo4j Inc. Allrights reserved 2023
24
By 2025, 50% of generative AI initiatives
will have improved reliability and
transparency by combining deep learning
foundation models with knowledge graphs
or other composite AI elements.
Technological Implications of Generative AI, August 2023
Impact Radar for GenAI (2024)
Neo4j Inc. Allrights reserved 2024
A Neo4j Knowledge Graph combined with LLM’s
obtains some unique improvements:
Accuracy - Obtain better answers compared
to plain vector searches
Specificity: domain specific, factual
knowledge on your subject
Explainability: Provide the user with more
reasoning on how the results were obtained.
Security: Role Based Access Control
Retrieval Augmented Generation
Evolving From RAG to GraphRAG
28.
We are notmaking this up…
Neo4j Inc. All rights reserved 2023
28
Neo4j Inc. Allrights reserved 2023
32
Data Science on Graphs: Graph Data Science…
Vector Search
Graph
Data Science
Knowledge
Graph
Bring the context of your connected data into
a format that other pipelines can ingest.
The Largest Catalog of
Graph Algorithms
Graph Vector Embeddings
for Machine Learning
Neo4j Inc. Allrights reserved 2023
34
GraphRAG with Neo4j
Find similar documents
and content
Identify entities
associated to content and
patterns
in connected data
Improve GenAI inferences
and insights. Discover new
relationships and entities
Unify vector search, knowledge graph and data science
capabilities to improve RAG quality and effectiveness
Vector Search
Graph
Data Science
Knowledge
Graph
Why a KGMatters in RFP GenAI App?
36
Challenges Outcomes
Time consuming to read previous
RFP across multiple repositories
Knowledge base to collect, store
and retrieve domain-specific
information
Repetitive and manual tasks to
synthesise the content
Drive efficient, accurate,
contextual and explainable way to
streamline RFP responses
Non-standard structure of RFP
making it difficult to do data
modelling
Flexible storage that’s adoptable
to the varying structure of an RFP
Neo4j Inc. All rights reserved 2024
37.
Anatomy of anRFP Document
37
AWS RFP
Intro Objectives Proposal
About the Company
Financial Result
Content Subsection 1
Subsection 2
Subsection 1.1
Content
Content Content
Content
Content
Content
Content
Content
Content
Content
Neo4j Inc. All rights reserved 2024
38.
Anatomy of anRFP Document
38
AWS RFP
Intro Objectives Proposal
About the Company
Financial Result
Content Subsection 1
Subsection 2
Subsection 1.1
Content
Content Content
Content
Content
Content
Content
Content
Content
Content
Neo4j Inc. All rights reserved 2024
Why Neo4j KGMatters in RFP GenAI App?
40
Challenges Outcomes
Time consuming to read previous
RFP across multiple repositories
Knowledge base to collect, store
and retrieve domain-specific
information
Repetitive and manual tasks to
synthesise the content
Drive efficient, accurate,
contextual and explainable way to
streamline RFP responses
Non-standard structure of RFP
making it difficult to do data
modelling
Flexible storage that’s adoptable
to the varying structure of an RFP
Neo4j Inc. All rights reserved 2024
41.
Why Neo4j KGMatters in RFP GenAI App?
41
Challenges Outcomes
Time consuming to read previous
RFP across multiple repositories
Knowledge base to collect, store
and retrieve domain-specific
information
Repetitive and manual tasks to
synthesise the content
Drive efficient, accurate,
contextual and explainable way to
streamline RFP responses
Non-standard structure of RFP
making it difficult to do data
modelling
Flexible storage that’s adoptable
to the varying structure of an RFP
Neo4j Inc. All rights reserved 2024
42.
AWS RFP
Intro ObjectivesProposal
Content
Chunk
Content
Chunk
Content
Chunk
Financial
Result
About the
Company
Content
Chunk
Content
Chunk
Content
Chunk
Subsection
2
Subsection
1
Content
Chunk
Content
Chunk
Subsection
1.1
Content
Chunk
Content
Chunk
Content
Chunk
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Knowledge Graph as the Knowledge Base
Document in a KG Knowledge Graph
42 Neo4j Inc. All rights reserved 2024
43.
Knowledge Graph asthe Knowledge Base
Document in a KG Knowledge Graph
43 Neo4j Inc. All rights reserved 2024
AWS RFP
Intro Objectives Proposal
Content
Chunk
Content
Chunk
Content
Chunk
Financial
Result
About the
Company
Content
Chunk
Content
Chunk
Content
Chunk
Subsection
2
Subsection
1
Content
Chunk
Content
Chunk
Subsection
1.1
Content
Chunk
Content
Chunk
Content
Chunk
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
44.
Why Neo4j KGMatters in RFP GenAI App?
44
Challenges Outcomes
Time consuming to read previous
RFP across multiple repositories
Knowledge base to collect, store
and retrieve domain-specific
information
Repetitive and manual tasks to
synthesise the content
Drive efficient, accurate,
contextual and explainable way to
streamline RFP responses
Non-standard structure of RFP
making it difficult to do data
modelling
Flexible storage that’s adoptable
to the varying structure of an RFP
Neo4j Inc. All rights reserved 2024
45.
Why Neo4j KGMatters in RFP GenAI App?
45
Challenges Outcomes
Time consuming to read previous
RFP across multiple repositories
Knowledge base to collect, store
and retrieve domain-specific
information
Repetitive and manual tasks to
synthesise the content
Drive efficient, accurate,
contextual and explainable way to
streamline RFP responses
Non-standard structure of RFP
making it difficult to do data
modelling
Flexible storage that’s adoptable
to the varying structure of an RFP
Neo4j Inc. All rights reserved 2024
46.
Accurate, Contextual andExplainable
46 Neo4j Inc. All rights reserved 2024
GenAI App
Who is the main respondent
of the AWS RFP?
47.
Accurate, Contextual andExplainable
47 Neo4j Inc. All rights reserved 2024
Who is the main
respondent of the
AWS RFP?
GenAI App
Embedding
Model
User
Question
48.
Accurate, Contextual andExplainable
48 Neo4j Inc. All rights reserved 2024
Who is the main
respondent of the
AWS RFP?
GenAI App
Embedding
Model
User
Question
Vector Embedding
49.
Accurate, Contextual andExplainable
49 Neo4j Inc. All rights reserved 2024
Who is the main
respondent of the
AWS RFP?
GenAI App
Embedding
Model
User
Question
Similarity Search using
Neo4j Vector Index
Vector Embedding
50.
Accurate, Contextual andExplainable
50 Neo4j Inc. All rights reserved 2024
AWS RFP
Intro Objectives Proposal
Content
Chunk
Content
Chunk
Content
Chunk
Financial
Result
About the
Company
Content
Chunk
Content
Chunk
Content
Chunk
Subsection
2
Subsection
1
Content
Chunk
Content
Chunk
Subsection
1.1
Content
Chunk
Content
Chunk
Content
Chunk
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Similarity Search using
Neo4j Vector Index
Response
from ABC
Company
Signed by
John
Smith
He’s the
General
Manager
51.
Accurate, Contextual andExplainable
51 Neo4j Inc. All rights reserved 2024
AWS RFP
Intro Objectives Proposal
Content
Chunk
Content
Chunk
Content
Chunk
Financial
Result
About the
Company
Content
Chunk
Content
Chunk
Content
Chunk
Subsection
2
Subsection
1
Content
Chunk
Content
Chunk
Subsection
1.1
Content
Chunk
Content
Chunk
Content
Chunk
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Contextual
Knowledge Retrieval
within Neo4j KG
Response
from ABC
Company
Signed by
John
Smith
He’s the
General
Manager
52.
Accurate, Contextual andExplainable
52 Neo4j Inc. All rights reserved 2024
AWS RFP
Intro Objectives Proposal
Content
Chunk
Content
Chunk
Content
Chunk
Financial
Result
About the
Company
Content
Chunk
Content
Chunk
Content
Chunk
Subsection
2
Subsection
1
Content
Chunk
Content
Chunk
Subsection
1.1
Content
Chunk
Content
Chunk
Content
Chunk
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Knowledge Retrieval
to aid in
Explainability
Response
from ABC
Company
Signed by
John
Smith
He’s the
General
Manager
53.
Accurate, Contextual andExplainable
53 Neo4j Inc. All rights reserved 2024
AWS RFP
Intro Objectives Proposal
Content
Chunk
Content
Chunk
Content
Chunk
Financial
Result
About the
Company
Content
Chunk
Content
Chunk
Content
Chunk
Subsection
2
Subsection
1
Content
Chunk
Content
Chunk
Subsection
1.1
Content
Chunk
Content
Chunk
Content
Chunk
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Fine Grained Access
Control to prevent
unwarranted Knowledge
Retrieval
Response
from ABC
Company
Signed by
John
Smith
He’s the
General
Manager
54.
Accurate, Contextual andExplainable
54 Neo4j Inc. All rights reserved 2024
Who is the main
respondent of the
AWS RFP?
GenAI App
Embedding
Model
User
Question
Similarity Search using
Neo4j Vector Index
Vector Embedding
55.
Accurate, Contextual andExplainable
55 Neo4j Inc. All rights reserved 2024
Who is the main
respondent of the
AWS RFP?
GenAI App
Embedding
Model
User
Question
Vector Embedding
AWS RFP
Intro Objectives Proposal
Content
Chunk
Content
Chunk
Content
Chunk
Financial
Result
About the
Company
Content
Chunk
Content
Chunk
Content
Chunk
Subsection
2
Subsection
1
Content
Chunk
Content
Chunk
Subsection
1.1
Content
Chunk
Content
Chunk
Content
Chunk
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Response
from ABC
Company
Signed by
John
Smith
He’s the
General
Manager
Similarity Result
Similarity Search using
Neo4j Vector Index
Bedrock
LLM
56.
Accurate, Contextual andExplainable
56 Neo4j Inc. All rights reserved 2024
Who is the main
respondent of the
AWS RFP?
GenAI App
Embedding
Model
User
Question
Vector Embedding
AWS RFP
Intro Objectives Proposal
Content
Chunk
Content
Chunk
Content
Chunk
Financial
Result
About the
Company
Content
Chunk
Content
Chunk
Content
Chunk
Subsection
2
Subsection
1
Content
Chunk
Content
Chunk
Subsection
1.1
Content
Chunk
Content
Chunk
Content
Chunk
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Response
from ABC
Company
Signed by
John
Smith
He’s the
General
Manager
Similarity Result
Similarity Search using
Neo4j Vector Index
The main respondent of
the AWS RFP is John
Smith, General Manager
XYZ Group representing
ABC Company (source:
RFP_AWS.pdf, page: 38)
AWS RFP
Intro Objectives Proposal
Content
Chunk
Content
Chunk
Content
Chunk
Financial
Result
About the
Company
Content
Chunk
Content
Chunk
Content
Chunk
Subsection
2
Subsection
1
Content
Chunk
Content
Chunk
Subsection
1.1
Content
Chunk
Content
Chunk
Content
Chunk
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Vector
Embedding
Response
from ABC
Company
Signed by
John
Smith
He’s the
General
Manager
Similarity +
Contextual Result
Bedrock
LLM
57.
How about Graph
DataScience?
57
Enrich the measure of relevancy
using graph algorithms.
● Page Rank to understand the
importance of parts of
documents
● Link Prediction to find hidden
relationships that further
contextualise the results
● Community Detection to group
related parts of documents for
more focused knowledge
retrieval
Neo4j Inc. All rights reserved 2024
GraphRAG enables youto..:
● To leverage structural
information across entities to
enable more precise and
comprehensive retrieval
● To perform advanced Graph
analytics to enhance retrieval
● To have an accurate conversation
with your data that is explainable
64