Generative AI with

Amazon Bedrock

Ganesh Gella
Director of Engineering, Amazon Lex &
Amazon Bedrock Agents, AWS

© 2023, Amazon Web Services, Inc. or its affiliates. All rights reserved. Amazon Confidential and Trademark.
 तोहय 4छवा वारय സുഖമാേണാ
!ర# ఎల' ఉ)*+ర# भवान ् कथमDस सभ कुशल मंगल
थआ
ु ढ़ा के( हाल ऐ আেপানাৰ (কেন

म बर यू ड(ग તમે ક&મ છો !ೕವ$ %ೇ'()ೕ*

आप कैसे ह0 .How are you?. ‫ﯿﺌﻦ آھﯿﻦ‬$ ‫ﺗﻮن‬

େକମିତ ି ଅଛ)ି, େକମିତ ି ଅଛ ‫آپ ﮐﯾﺳﮯ ﮨو‬ सेतलेका मेनमा?

तू कसा आहे स 4तमीलाई कBतो छ
আপুিন (কিন আেন? আপিন (কমন আেছন ਤੁਸੀ ਿਕਵ) ਹੋ

எ"ப$ இ'(கிற,-க. कएम डबबरH बे गे

 TRADITIONAL ML MODELS FOUNDATION MODELS

Train Deploy Tasks Pretrain Adapt Tasks

Text generation Text generation

Summarization Summarization

Labeled data ML models Info extraction Unlabeled Info extraction

FM
data
Q&A Q&A

Chatbot Chatbot
 Essentials for building a
generative AI application

Purpose-built Access to a variety Private environment Easy-to-use tools to build

ML infrastructure of foundation models to leverage your data and deploy applications
 2010 2020 2023

CG1 G2 P2 G3 P3 G4 P4 G5 G5g P5
NVIDIA Tesla NVIDIA GRID NVIDIA NVIDIA NVIDIA V100 NVIDIA T4 NVIDIA A100 NVIDIA A10G NVIDIA T4G NVIDIA H100
M2050 “Fermi” GK104 “Kepler” K80 Tesla M60 Tensor Core Tensor Core Tensor Core Tensor Core Tensor Core Tensor Core
GPUs GPUs GPUs GPUs GPUs GPUs GPUs GPUs GPUs GPUs
 Innovating at the silicon level

AWS AWS
 Inferentia2

HIGHER THROUGHPUT LOWER LATENCY

 Essentials for building a
generative AI application

Purpose-built Access to a variety Private environment Easy-to-use tools to build

ML infrastructure of foundation models to leverage your data and deploy applications
No one model
 Amazon
Broad choice of models

Jurassic Amazon Titan Claude Command + Embed Llama 2 Stable Diffusion

Contextual answers, Text summarization, Summarization, complex Text generation, search, Q&A and reading High-quality images and
summarization, generation, Q&A, search reasoning, writing, coding classification comprehension art
paraphrasing
 NEW

Now Available in Amazon Bedrock

Claude 2.1

Industry-leading 200K 2x reduction in model Reduces cost of prompts and

token context window hallucination rate completions on Bedrock by 25%
 NEW

Now Available in Amazon Bedrock

Claude 2.1 Llama 2 70B

Fine-tuned for chat use cases Supports large-scale tasks

 NEW NEW

Amazon Amazon
Summarization, copywriting, Open-ended text generation,
and ideal for fine-tuning conversational chat, and RAG support

GENERALLY AVAILABLE GENERALLY AVAILABLE

 .NEW.

Generate studio-quality images using natural

Amazon Titan language prompts

Image Generator Customize images with proprietary data

to match your brand style

Generate realistic, studio-quality images

Higher scores for text-to-image alignment
compared to several leading models
AVAILABLE IN PREVIEW TODAY
 watermarks
support responsible AI
IMAGE GENERATION

Show me an image of an iguana

OUTPAINTING

Image of an iguana in a rainforest

IMAGE VARIATION

orange iguana facing right in a rainforest

Vector embeddings are numerical representations of text

You[[ may return or 0.0057839

0.00526205 exchange your order within... 30
0.00856714
days 0.00083171
of when the order was shipped.
-0.00273974 Full refunds
-0.01069805] [ will
be processed back0.01325288
0.02154857 to the original payment method.
0.00202334 ... -
After 30 days,
0.01588322 we offer a store
-0.00311089 credit only.[
0.01122954]
Customers will be0.00205388
0.00033921 responsible for the cost to...ship
0.00069172
returns or-0.01132295
0.00243699 exchanges after 30 days. [
0.00119721]
0.00333424 0.00254636 0.0116470
 Cat

Kitten Old Canine

Vectors are critical for

customizing generative Felin
e

AI applications
Puppy Dog

Young
 Supports semantic search, text retrieval,
Amazon Titan
and clustering
Text Embeddings
Supports a context length of up to 8k
tokens
Translates text into numerical
representations Works with 25+ languages
 WITHOUT VECTOR EMBEDDINGS WITH VECTOR EMBEDDINGS

Golf Shoes Bag Golf Shoes

Golf Shoes Bag golf

Golf Shoes Bag

Golf Shoe Case Golf Shoes

bright color golf shoes

Stretcher for Golf Shoes Golf Shoes

Golf
Golf Shoes
Shoes

Plastic Golf Tees. Golf Shoes

IMAGE SEARCH

Upload your image

Show me what works well with my sofa

 .NEW.

Accepts text, image, or a combination

Amazon Titan of text-image to generate embeddings

Multimodal Adapts to unique and proprietary data

Embeddings Built-in mitigation to help reduce

biased search results
Search, recommendation, and personalization

GENERALLY AVAILABLE
 Amazon Titan Foundation Models

TITAN TEXT TITAN TEXT TITAN TEXT TITAN MULTIMODAL TITAN IMAGE
EMBEDDINGS LITE EXPRESS EMBEDDINGS GENERATOR

Translates text into Summarization, Open-ended text Search, Generate realistic,

numerical representations copywriting, generation, recommendation, studio-quality
fine-tuning conversational chat, personalization images
RAG support
 Amazon
Broad choice of models

Jurassic-2 Ultra Titan Text Embeddings Claude 2 Command + Embed Llama 2 Stable Diffusion XL1.0
Jurassic-2 Mid Titan Multimodal Embeddings Claude 2.1 Cohere Command Light Llama 2 13B
Titan Text Lite Claude Instant Cohere Embed English Llama 2 70B
Titan Text Express Cohere Embed Multilingual
Titan Image Generator
More than are using Amazon Bedrock
 Essentials for building a
generative AI application

Purpose-built Access to a variety Private environment Easy-to-use tools to build

ML infrastructure of foundation models to leverage your data and deploy applications
for generative AI applications
Adapt models for FOUNDATION
MODEL

your use case with

fine tuning
FINE-TUNING LABELED
DATA
 FOUNDATION
MODEL Update your models
through continued
pre-training
FINE-TUNING UNLABELED
DATA
 Fine tune additional models
in Amazon Bedrock
COMING SOON
 QUALITY

.NEW.

Model evaluation on TRADE OFF

Amazon Bedrock
Evaluate, compare, and select the best
foundation model for your use case
PREVIEW

COST LATENCY
 Is Finetuning the only option to make
Large Language Models work with my Data ?
 CONVERT
data into embeddings

STORE
integrations with vector databases
Implementing RAG can
RETRIEVE
find relevant results of from vector database
based on user’s query

AUGMENT
add above results along with user’s query to
augment the prompt

GENERATE
instruct LLMs to generate responses based on
contextual data
Knowledge Bases for Automatically converts text documents into embeddings

Stores embeddings in your vector database

Fully-managed native support for

Retrieves embeddings and augments prompts
retrieval augmented generation
 What if I have more dynamic data or
complex tasks for my Application needs ?
Agents for
Amazon Bedrock
Enable generative AI applications to
complete tasks in just a few clicks
 Enabling foundation models to execute tasks

1 2 3 4

SELECT YOUR PROVIDE BASIC DEVELOPER SPECIFIES SELECT RELEVANT

FOUNDATION MODEL INSTRUCTIONS LAMBDA FUNCTIONS DATA SOURCES
 Essentials for building a
generative AI application

Purpose-built Access to a variety Private environment Easy-to-use tools to build

ML infrastructure of foundation models to leverage your data and deploy applications
 Amazon Q Amazon Q in Amazon Q in Amazon
Amazon QuickSight Amazon Connect CodeWhisperer

Generative Amazon Bedrock

AI Stack
Guardrails Agents Customization capabilities

GPUs Trainium Inferentia SageMaker

UltraClusters EFA EC2 Capacity Blocks Nitro Neuron

 Amazon Q accelerates productivity
AMAZON Q is AMAZON Q

AMAZON Q is AMAZON Q
Everything you need to accelerate your
Go Create
with AWS