AWS Summit Seoul 2023 | 생성 AI 모델의 임베딩 벡터를 이용한 서버리스 추천 검색 구현하기

© 2023, Amazon Web Services, Inc. or its affiliates. All rights reserved.
S E O U L | M A Y 4 , 2 0 2 3

생성 AI 모델의 임베딩벡터를 이용한
서버리스 추천/검색 구현하기
이상현
CEO
Mirror Co., Ltd.

• AWS Serverless Hero
• CEO, Mirror
• CTO,
• Tech Lead,
/ Kurt Lee
breath103@gmail.com
https://github.com/breath103

1. AI /
2. AI / +
3. Serverless
Agenda

function cosineSimilarity(a: Array<Number>, b: Array<Number>) {
return dotProduct(a, b) / (length(a) * length(b))
}
function dotProduct(a: Array<Number>, b: Array<Number>) {
return zip(a, b).map((pair) => pair[0] * pair[1]).sum()
}
function length(a: Array<Number>) {
return sqrt(a.sum(s => Math.pow(s, 2)));
}
Disclaimer
• AI / ( ). Software Engineer
• .
•

1. AI

ML Model

Deep Neural Network

…

Convolutional Neural Network

/ / RNN

Transformer

..
Activation Function,
Learning Rate,
Parameter count
…
variation

,
( )

1. Encoder
Input “ ”
(Human readable Input) → (Latent Vector)
Decoder
“ ” Output
(Latent Vector) → (Human readable Output)

ex) CNN Encoder

1. “ ” Embedding Vector
2. “ ”
3. .

Vector “ ” ?

Not a joke

[0.10, 0.95]
[0.60, 0.60]
[0.95, 0.10]
[0.10, 0.10]

1. , “ (feature)”
/
Latent Vector Space
( = Latent Vector Space)
Latent Vector =

,
3 .
“ ” , Vector
=

PCA / T-SNE ,
A.I. Experiments: Visualizing High-Dimensional Space
https://www.youtube.com/watch?v=wvsE8jm1GzE

Latent space / Latent Feature Space / Embedding Space

function cosineSimilarity(a: Array<Number>, b: Array<Number>) {
return dotProduct(a, b) / (length(a) * length(b))
}
function dotProduct(a: Array<Number>, b: Array<Number>) {
return zip(a, b).map((pair) => pair[0] * pair[1]).sum()
}
function length(a: Array<Number>) {
return sqrt(a.sum(s => Math.pow(s, 2)));
}

2. Embedding Vector /

“ Vector ”
/

Matrix Factorization / Collaborative Filtering
/ Latent Vector

?
1. vector
2. vector
3. vector
vector
1. vector
2. vector
3. vector

ML ?

1. “ ” .
( inference )
• , ,
• ex) “ ” → “ classification ”
• ,
• “ ” , .
• → GPT / Llama / BERT ,
Large Language Model
• → ImageNet (ResNet,
SAM…)

2. embedding .
• PyTorch / Tensorflow raw-level access :
.
embeddings .
• Huggingface , :
Huggingface transformers ,
.
• OpenAI , API SaaS :
text Array<Float> API .

🤗 Huggingface

🤗 Huggingface + Local Machine (BART)

🤗 Huggingface + SageMaker Inference

3. Context

3. Serverless

Serverless
• Computing Platform
• ( , …)
• (Memory, CPU, Persistency…)
• , , …
• Serverless ? ( X )
• Memory / CPU / Local storage ? ( O )
• Highly distributed computing system ? ( O )

)

• , emoji
Unicode ,
,
.
“🏠️”: “ , , ” — “ ” ?
“🏖️”: “ , ” — “ ” ?
“💰”: “ , , ” — “ ” ?
“🤖”: “ ” — “AI” ?
🎉🎉 Emoji 🎉🎉

1. Emoji + Vector (GPT-3.5 )
emoji_openai_embeddings.json (36 MB)

2. AWS Lambda
// Global context
const allEmojis: Array<{
emoji: string,
embeddings: number[]
}> = require("./emoji_openai_embeddings.json");
async function handler(event: { label: string }) {
// openAI embedding vector
const embedding: [number] = _.first((await OpenAIEmbeddings.getEmbeddings([event.label])));
return _.chain(allEmojis) // ,
.map((candidate) => {
//
return {
emoji: candidate.emoji,
distance: Math.sqrt(
_.chain(candidate.embeddings)
.sumBy((value, index) => {
Math.pow(candidate.embeddings[index] - embedding[index], 2)
})
) //
};
})
.sortBy((r) => r.distance)
.head(4) // 4
.map((r) => r.emoji)
}

Demo

Application
• ,
100%
• Q&A : “ ” =>
“Q: ”

• ?
Emoji embedding = 36MB
embedding ,
?
Serverless ?

Vector Database
https://www.pinecone.io/learn/vector-database/

Redis Vector Database
https://redis.io/docs/stack/search/reference/vectors/
Postgres pgVector
https://redis.io/docs/stack/search/reference/vectors/

SaaS

Batch Computing / ?
서비스 종류
Fully
Serverless
Managed
데이터 수집 Kinesis Kafka
대용량 데이터 저장
S3 / Redshift
Serverless
Redshift /
ElasticSearch…
데이터 처리 Lambda Amazon EMR
실시간 데이터 저장 DynamoDB Redis / RDS

ML + Serverless = 🎉🎉🎉

.

AWS Summit Seoul 2023 | 생성 AI 모델의 임베딩 벡터를 이용한 서버리스 추천 검색 구현하기

More Related Content

What's hot

More from Amazon Web Services Korea

AWS Summit Seoul 2023 | 생성 AI 모델의 임베딩 벡터를 이용한 서버리스 추천 검색 구현하기