© 2023, Amazon Web Services, Inc. or its affiliates.

All rights
reserved.
AIM307

Retrieval Augmented
Generation with Amazon
Bedrock
Rupinder Grewal Clay Elmore
(he/him) (he/him)
Sr. Specialist Solutions Architect Specialist Solutions Architect -
– ML AI/M L
AWS AWS

© 2023, Amazon Web Services, Inc. or its affiliates. All rights

reserved.© 2023, Amazon Web Services, Inc. or its affiliates. All
reserved.
rights reserved.
Agen
da
01 Introduction to Retrieval Augmented Generation
(RAG)
02 Workshop setup

03 Hands-on lab

04 Conclusions and next steps

© 2023, Amazon Web Services, Inc. or its affiliates. All rights

reserved.
Why RAG is a common solution in
generative AI apps

Challenges with RAG as a Components

LLM s solution • Retrieval mechanisms
• LLM training • Provides context to • Model choices
data is LLMs outside of
incomplete training data • Orchestration
frameworks
• Model • Reduces factual
hallucination inaccuracy
• Model • Limited complexity
customization
© 2023, Amazon Web Services, Inc. or its affiliates. All rights
reserved.
How RAG
works
User
input
Tex Use
Prompt
augmentati
Large
language
Respon
r se
t
generati on model

workflo
on Embeddin
gs
Contex
t
w model

Embeddin 0.89
0.17
-0.02 -0.53 0.95 -
0.38

Dat Semant
ingesti
a ic
search
on
workflo Vector Embeddings Document New
w store model store data
© 2023, Amazon Web Services, Inc. or its affiliates. All rights
reserved.
Amazon
Bedrock

Amazon Choose Use as is or Send Receive

an FM customize prompt response
Bedrock
Use the Fine-tune FMs as Use Bedrock API to Receive
Build generative AI playground to needed. Bedrock send your model
applications using experiment with will automatically prompts to the response in
foundation models FMs and select deploy the FM for model your
(FMs) through a the one that inference application
serverless suits your
API service needs

© 2023, Amazon Web Services, Inc. or its affiliates. All rights

reserved.
Titan
Embeddings
Amazon NE
W
Highlig
Titan hts
Embedding • Titan Embeddings offers fast,
s cost-effective, high-performance, and
accurate embeddings in 25
V2.0 languages.

Translates text inputs (words, phrases) into • Optimized for text retrieval tasks,
numerical representations (embeddings). semantic
Comparing embeddings produces more similarity, and clustering.
relevant and contextual responses than word
matching. • Applications of this model includes
semantic search and personalization.

Max tokens: 8,000

Output vectors: 1,536
Language: Multilingual (25 languages)

Model ID:Amazon
© 2023, amazon.titan-embed-text-v1
Web Services, Inc. or its affiliates. All rights
reserved.
Anthropic
Claude
Claud Highlig
e hts
Anthropic offers the Claude family of large • Long context window (100k)
language models purpose-built for allows for large amounts of text
conversations, summarization, Q&A, workflow to be processed at once.
automation, coding, and more. Claude can
also take direction on personality, tone, and • Available in two sizes to help
behavior. choose the right-sized model
based on latency, accuracy, and
cost considerations.
Max tokens: 100,000
Language: English and multiple other • Early customers report that
languages Claude is much less likely to
Model IDs: produce harmful outputs, easier
- anthropic.claude-instant-v1 to converse with, and more
- anthropic.claude-v2 steerable .

© 2023, Amazon Web Services, Inc. or its affiliates. All rights

reserved.
Let’s get
started!

© 2023, Amazon Web Services, Inc. or its affiliates. All rights

reserved.
 Workshop Link

http://bit.ly/3T8B
LSi

© 2023, Amazon Web Services, Inc. or its affiliates. All rights

reserved.
 Ne

Knowledge
w

base for
RAG
Connect FMs to data
sources, including vector
engine for Amazon
OpenSearch Serverless,
Pinecone, and Redis
Enterprise Cloud

Enable automatic data

source detection

Provide source attribution

© 2023, Amazon Web Services, Inc. or its affiliates. All rights

1
reserved. 1
Agents enable generative AI
Ne
w

applications to complete tasks

in just a few clicks

1 2 3 4

Select your Provide Select Developer

foundation basic relevant specifies
model instructio data Lambda
ns sources functions
Breaks down and orchestrates tasks
Securely accesses and retrieves company
data Takes action by invoking API calls on
your behalf Provides fully managed
infrastructure support
© 2023, Amazon Web Services, Inc. or its affiliates. All rights
reserved.
Further
considerations
Prompt RAG Vector
engineering orchestration databases

© 2023, Amazon Web Services, Inc. or its affiliates. All rights

reserved.
Continuing your Amazon Bedrock
journey
BePdrroomckptReAnGgiwnoe BeRdAroGcokrcohdeestsra VKencotworleddagtae
rekrisnhgop mtiopnles bbaassees

© 2023, Amazon Web Services, Inc. or its affiliates. All rights

reserved.
Thank
Thank you! Please complete the
Please complete the
session
survey in the mobile app
session

you!
Rupinder Grewal Clay Elmore
survey in the mobile app

[email protected] [email protected]
om om

© 2023, Amazon Web Services, Inc. or its affiliates. All rights

reserved.© 2023, Amazon Web Services, Inc. or its affiliates. All
reserved.
rights reserved.