to generative AI “A golden retriever wearing glasses and a hat in a portrait painting” > Question: What is generative AI? Generative AI is a branch of AI that focuses on creating new data. It is a subset of machine learning. The goal of generative AI is to create new data that is similar to the data that was used to train the model. Chat:
does generative AI work? Foundation model Text input Output Text generation model (also known as large language model) Image generation model Video Audio Code generation model “Summarize this article …….” [Text] “………..” “a photo of an astronaut riding a horse on mars” “A young couple walking in rain.” “Children singing nature songs” “Write Python code to sort array …” [Image] [Video] [Audio] [Code]
AI is powered by Foundation Models (FMs) Text generation Summarization Information extraction Q&A Chatbot Pretrain Adapt Tasks Unlabeled data Foundation model Text generation Summarization Information extraction Q&A Chatbot Train Deploy Tasks ML models … … … … Labeled data … … … …
AI and Foundation Models Generative AI is a type of AI that can create new content and ideas, including conversations, stories, images, videos, and music. Like all AI, generative AI is powered by ML models—very large models that are pre-trained on vast amounts of data and commonly referred to as Foundation Models (FMs). NLP – Natural Language Processing CV – Computer Vision LLM – Large Language Model FM – Foundation Model LLM CV NLP NLP – Natural Language Processing CV – Computer Vision LLM – Large Language Model FM – Foundation Model FM
Training and Inference How do I access Foundation Models on AWS? Amazon SageMaker JumpStart Amazon Bedrock Prompt / text embeddings API Layer SageMaker Endpoint Foundation Models SageMaker Jumpstart Model hub, deploy, fine-tune Accelerated Computing Trn1(n), Inf2, P4d, P5 Fine-tune API Layer Amazon Bedrock Foundation Models Prompt / text embeddings Fine-tune
Search for topics or problem types, and get relevant results across all content • Browse by content type to explore solutions, models, example notebooks, blogs, and video tutorials Browse and search SageMaker JumpStart content
Titan Summarization, image and text generation and search, Q&A Command & Embed Text generation, search, classification Jurassic-2 Contextual answers, summarization, paraphrasing Claude 3, Claude 2.1 & Claude Instant Multimodal reasoning, vision capabilities, translation, summarization, writing, coding Stable Diffusion XL 1.0 High-quality images and art Llama 2 Dialogue use cases and language tasks Mistral 7B & Mixtral 8x7B Text summarization, Q&A, Text classification, Text completion, code generation Mistral AI Amazon Bedrock supports leading foundation models
API to build with generative AI • Pass the model ID, type of content, and body of the request • Body includes the prompt and execution parameters • Returns model response and metadata • Handles text-to-text, text-to-image, image-to-image, and more • Supports current and future Amazon Titan models, third-party models, and even fine-tuned models
core API: Streaming responses response = bedrock.invoke_model_with_response_stream( modelId = model_id, body = body) stream = response.get('body') if stream: for event in stream: chunk = event.get('chunk’) if chunk: print(json.loads(chunk.get('bytes').decode())) Amazon Bedrock chunk 1 chunk N … invoke with stream 1 2 • Users can start reading the response as soon as the first chunk is available • For more information, see Amazon Bedrock Developer Guide
with LangChain • New LLM and embeddings class for Amazon Bedrock • Includes code for using the LLM class in a conversation chain • Includes code for creating an embedding from text from langchain import Bedrock from langchain.embeddings import BedrockEmbeddings llm = Bedrock() print(llm(“what is generative AI?”)) pip install langchain
Training and Inference Two ways to access Foundation Models on AWS Amazon SageMaker JumpStart • Machine learning (ML) hub with foundation models (public and proprietary) • Built-in algorithms, and prebuilt ML solutions • Deploy FM as SageMaker Endpoint (hosting) • Fine-tuning leverages SageMaker Training jobs • Choose SageMaker managed accelerated computing instance Prompt / text embeddings API Layer SageMaker Endpoint Foundation Models SageMaker Jumpstart Model hub, deploy, fine-tune Accelerated Computing Trn1(n), Inf2, P4d, P5 Fine-tune Amazon Bedrock • The easiest way to build and scale generative AI applications with foundation models (FMs) • Access directly or fine-tune foundation model using API • Serverless • Foundation model providers – Amazon, Anthropic, AI21, Stability, and many more. API Layer Amazon Bedrock Prompt / text embedding Fine-tune Foundation Models
need to customize a Foundation Model, why? 조선왕조실록에 기록 된 세종대왕 맥북프로 던짐 사건에 대해 알려줘 Hallucination Current Knowledge 2024년 3월 20일 Amazon 주식의 종가를 알려줘 Closed-domain Knowledge 나는 Amazon 직원인데, 고객 문의에 대한 응대 매뉴얼을 알려줘 아쉽게도, 현재 제가 직접적인 매뉴얼을 제공할 수 없습니다. Amazon에 근무하는 경우, 내부 교육 자료, 팀 리더 또는 교육 담당자와 상담하여 정확한 정보와 지침을 얻을 수 있습니다. 세종대왕이 맥북 프로를 던지는 사건은 조선왕조실록에 기록되어 있습니다. 이 사건은 1443년(세종 25년) 세종이 집현전 학자들과 함께 궁중 후원에서 열린 연회에서 벌어진 일입니다. 세종은 연회 도중에 갑자기 맥북 프로를 바닥에 던져버렸습니다. 죄송합니다만, 제 현재 정보는 2022년 3월 기준이며 최신 주식 가격 정보는 제가 알 수 없습니다. 최신 정보를 확인하려면 실시간 주식 시세 사이트나 금융 뉴스를 참고하시길 바랍니다.
need to customize a Foundation Model, why? • Specific Task • Closed-domain knowledge • Current Knowledge • Improving the performance/quality • Reduce likelihood of hallucinations Foundation Model Common Crawl Wikipedia Initial Pretraining Prompt (Question) Response (Answer) Which FMs are supported by Amazon Bedrock? Radio X FM & TV
Engineering, new way of using ML! Task Description Example Input 1 Example Input 2 Example Input 3 Output Indicator Generated Response Input Output Prompt LLM
review sentiment classifier. Review: "I loved this movie!" This review is positive. Review: "I am not sure, I think the movie was fine.” This review is neutral. Review: "This movie was a waste of time and money" This review is negative. Review: "I really had fun watching this movie” This review is Positive Input Output Model Task Description Examples Output indicator Prompt Engineering, new way of using ML!
to do Prompt Augmentation? 맥북 프로(MacBook Pro)는 애플 매킨토시 노트북 컴퓨터 라인업으로, 2006년 1월 10일 맥월드 엑스포에서 처음 공개되었다. 세종은 조선의 제4대 왕이다. 재위 기간은 1418~1450년이며, 1418년 6월에 왕세자에 책봉되었다가 8월에 태종의 양위를 받아 즉위했다. Prompt Context Generated Response
does RAG work? User Query: HDL 콜레스테롤이 뭐지? Search Relevant Information Medical Knowledge Sources Relevant Information for Enhanced Context: HDL 콜레스테롤은 쉽게 얘기하면 좋은 콜레스테롤 이라고 말할 수 있습니다. HDL 콜레스테롤 수치가 낮다면 심혈관계 위험인자가 되기 때문에 HDL 콜레스테롤 수치를 높여야 합니다. Generated Text Response: HDL 콜레스테롤은 혈액 속 남아있는 콜레스테롤을 간으로 운반하여 배설되게 하기 때문에, 혈관을 청소하는 ‘좋은 콜레스테롤'로 불려요. 1 3 4 5 2 Query: HDL 콜레스테롤이 뭐지? Enhanced Context: HDL 콜레스테롤은 쉽게 얘기하면 좋은 콜레스테롤 이라고 말할 수 있습니다. HDL 콜레스테롤 수치가 낮다면 심혈관계 위험인자가 되기 때문에 HDL 콜레스테롤 수치를 높여야 합니다. API Layer Foundation Models
to capture Semantic Relationships between Words? Male-Female Verb tense King Queen Man Woman walking walked swiming swam Spain Madrid Korea Seoul Japan Tokyo China Beijing Russia Moscow U.S.A Washington D.C Country-Capital
is a vector embedding? • A numerical representation of words or sentences, used in NLP • NLP models can easily perform tasks such as querying, classification, and applying machine learning algorithms on textual data “I am at AWS Unicorn Day 2024“ [0.743, 0.720, -0.325, 0.195, 0.835, -0.945] n-dimensional vector Magnitude Direction
similarity search Similar vectors = similar meaning and context Deliver similarity search results based on vector proximity by comparing vector distances Your favorite song Song with a high chance of being liked Amazon Music: Recommendations
(L2) Useful for counts/measurements recommendation systems Dot product Useful for collaborative filtering Cosine similarity Useful for semantic search and document classification Similarity Metrics for Vector search
using Vector Database D A T A I N G E S T I O N W O R K F L O W Prompt + Query Search Relevant Information Knowledge Sources Query LLM Endpoint Generated Text Response 1 2 Relevant Information for Enhanced Context 3 5 4 Enhanced Context Query Prompt + +
– Titan Embeddings “This is just some input I want to get embeddings for “ [0.752, 0.271, -0.325, 0.195, 0.837, -0.945] n-dimensional vector Amazon Bedrock
Chunking Strategies for RAG Amazon Bedrock 및 Amazon Titan을 통해 제공될 예정인 생성형 AI(GenAI; Generative AI)와 대규모 언어 모델(LLM; Large Language Model)은 개발자와 기업 이 자연어 처리 및 이해와 관련된 전통적으 로 복잡했던 문제를 해결하는 방식을 변화 시키고 있습니다. LLM이 제공하는 몇 가지 이점으로는 고객 서비스 애플리케이션에서 더 강력하고 매력적인 대화형 AI 환경을 구 축하고, 더 직관적이고 정확한 응답 결과를 통해 직원 생산성을 향상시킨다는 것입니 다. 그러나 이러한 유스케이스를 위해서는 대 화형 경험을 구현하는 GenAI 애플리케이션 이 회사 데이터에 대한 응답을 제한하여 모 델 환각(hallucinations, 잘못된 진술)을 완 화하고 최종 사용자의 콘텐츠 액세스 권한 에 따라 응답 결과를 필터링하는 두 가지 주 요 기준을 충족하는 것이 중요합니다. Sentence splitting Amazon Bedrock 및 Amazon Titan을 통해 제공될 예정인 생성형 AI(GenAI; Generative AI)와 대규모 언어 모델(LLM; Large Language Model)은 개발자와 기업 이 자연어 처리 및 이해와 관련된 전통적으 로 복잡했던 문제를 해결하는 방식을 변화 시키고 있습니다. LLM이 제공하는 몇 가지 이점으로는 고객 서비스 애플리케이션에서 더 강력하고 매력적인 대화형 AI 환경을 구 축하고, 더 직관적이고 정확한 응답 결과를 통해 직원 생산성을 향상시킨다는 것입니 다. 그러나 이러한 유스케이스를 위해서는 대 화형 경험을 구현하는 GenAI 애플리케이션 이 회사 데이터에 대한 응답을 제한하여 모 델 환각(hallucinations, 잘못된 진술)을 완 화하고 최종 사용자의 콘텐츠 액세스 권한 에 따라 응답 결과를 필터링하는 두 가지 주 요 기준을 충족하는 것이 중요합니다. Page splitting Page overlapping Amazon Bedrock 및 Amazon Titan을 통해 제공될 예정인 생성형 AI(GenAI; Generative AI)와 대규모 언어 모델(LLM; Large Language Model)은 개발자와 기업 이 자연어 처리 및 이해와 관련된 전통적으 로 복잡했던 문제를 해결하는 방식을 변화 시키고 있습니다. LLM이 제공하는 몇 가지 이점으로는 고객 서비스 애플리케이션에서 더 강력하고 매력적인 대화형 AI 환경을 구 축하고, 더 직관적이고 정확한 응답 결과를 통해 직원 생산성을 향상시킨다는 것입니 다. 그러나 이러한 유스케이스를 위해서는 대 화형 경험을 구현하는 GenAI 애플리케이션 이 회사 데이터에 대한 응답을 제한하여 모 델 환각(hallucinations, 잘못된 진술)을 완 화하고 최종 사용자의 콘텐츠 액세스 권한 에 따라 응답 결과를 필터링하는 두 가지 주 요 기준을 충족하는 것이 중요합니다.
Training and Inference Two ways to access Foundation Models on AWS Amazon SageMaker JumpStart Amazon Bedrock Prompt / text embeddings API Layer SageMaker Endpoint Foundation Models SageMaker Jumpstart Model hub, deploy, fine-tune Accelerated Computing Trn1(n), Inf2, P4d, P5 Fine-tune API Layer Amazon Bedrock Foundation Models Prompt / text embeddings Fine-tune
Generative AI application with RAG approach Amazon Kendra Index with Data Ingested User 1 6 Generative AI Application Prompt + Query 1 Query 2 Relevant Information for Enhanced Context 3 Generated Text Response 5 4 Prompt + Query + Enhanced Context Response from LLM to User 6 5 4 Amazon SageMaker (LLM for Text Generation) 2 3 (source: https://tinyurl.com/y3b7fvfh)
when it comes to implementing RAG, there are challenges… Creating vector embeddings for large volumes of data Orchestration Managing multiple data sources Scaling retrieval mechanism Coding effort Incremental updates to vector store
N S W E R Q U E S T I O N V E C T O R D A T A B A S E B E D R O C K F O U N D A T I O N M O D E L Knowledge bases for Amazon Bedrock U S E R E T R I E V A L A U G M E N T E D G E N E R A T I O N ( R A G ) Connect FMs to data sources including vector engine for Amazon OpenSearch Serverless, Amazon Aurora PostgreSQL, Pinecone, and Redis Enterprise Cloud Enable automatic data source detection Easily retrieve relevant data and augment prompt Provide citations
Language Model Prompt augmentation Response Embeddings model Data source Vector store Embeddings model Embedding User User Input Context -0.02 0.89 -0.38 -0.53 0.95 0.17 Text Generation Workflow Data Ingestion Workflow Semantic search Document chunks End-to-End RAG Workflow
using Knowledge Bases for Amazon Bedrock D A T A I N G E S T I O N W O R K F L O W Vector embeddings Document Chunks Embeddings model Data source Vector store Knowledge Bases for Amazon Bedrock
Ingestion Workflow K N O W L E D G E B A S E S F O R A M A Z O N B E D R O C K • Choose your data source (Amazon S3) • Support for incremental updates • Multiple data file formats supported • Choose your chunking strategy • Fixed chunks • No chunking • Default (200 tokens) • Choose your embedding model • Amazon Titan • Cohere • Choose your vector store • Open Search Serverless • Amazon Aurora PostgreSQL • Pinecone • Redis Enterprise Embeddings model New data Fully managed data ingestion workflow Document chunks Data source Vector store
managed data ingestion K N O W L E D G E B A S E S F O R A M A Z O N B E D R O C K • Choose your data source (Amazon S3) • Support for incremental updates • Multiple data file formats supported • Choose your chunking strategy • Fixed chunks • No chunking • Default (200 tokens) • Choose your embedding model • Amazon Titan • Cohere • Choose your vector store • Open Search Serverless • Amazon Aurora • Pinecone • Redis Enterprise Embeddings model New data Fully managed data ingestion workflow Document chunks Data source Vector store Automated and fully managed data ingestion using Knowledge Bases for Amazon Bedrock
and Generate Large Language Model Prompt augmentation Response Knowledge Bases for Amazon Bedrock Embeddings model Embedding User User Input Context -0.02 0.89 -0.38 -0.53 0.95 0.17 Text Generation Workflow Semantic search K N O W L E D G E B A S E S F O R A M A Z O N B E D R O C K
API K N O W L E D G E B A S E S F O R A M A Z O N B E D R O C K Response User User Input RetrieveAndGenerate API Generate query embedding Retrieve similar documents from knowledge bases Augment query with retrieved documents Generate response from LLM User query Generated response Fully managed RAG
RAG workflows using Retrieve API K N O W L E D G E B A S E S F O R A M A Z O N B E D R O C K Large Language Model Prompt augmentation Response User User Input Context Customized RAG workflow Retrieve API Generate query embedding Retrieve similar documents from knowledge bases User query Retrieved documents
use cases Improved content quality E.g., helps in reducing hallucinations and connecting with recent knowledge including enterprise data Contextual chatbots and question answering E.g., enhance chatbot capabilities by integrating with real-time data Personalized search E.g., searching based on user previous search history and persona Real-time data summarization E.g., retrieving and summarizing transactional data from databases, or API calls
Knowledge Bases for Amazon Bedrock RAG concepts Vector Database A N S W E R Q U E S T I O N V E C T O R D A T A B A S E B E D R O C K F O U N D A T I O N M O D E L Amazon SageMaker Amazon Bedrock Gen AI
[1] Generative AI on Amazon SageMaker Workshop https://catalog.us-east-1.prod.workshops.aws/workshops/972fd252-36e5-4eed-8608-743e84957f8e/en-US [2] Building with Amazon Bedrock and LangChain Workshop https://catalog.workshops.aws/building-with-amazon-bedrock/en-US [3] Building a Chatbot using Amazon Bedrock and Kendra Workshop https://catalog.us-east-1.prod.workshops.aws/workshops/27eb3134-4f33-4689-bb73-269e4273947a/en-US [4] RAG Architecture Recipes with sample codes https://www.awsstartup.io/startup-recipe/architecture-recipes#generative-ai
ksmin23
pospome
bkuhlmann
seyfoyun
zsmb
schroneko
youkidearitai
okashoi
mackey0225
hanetsuki
badmintoncryer
yossydev
robhawkes
keavy
cherdarchuk
pauljervisheath
samlambert
vanstee
kneath
smashingmag
lemiorhan
lauravandoore
holman
sugarenia
