Git Product home page Git Product logo

rag-tutorial's Introduction

Welcome to my repo! This is Inhyeok Yoo ๐Ÿ‘‹

  • 2011 - 2016: Bachelor's degree in industrial engineering, Kangwon National University
  • 2017 - 2019: master's degree in industrial engineering, Inha university
  • 2020 - 2021: Smart City Institute at VAIV company, Sejong special city, Republic of South Korea.
  • 2021 - Present: AI solution team at Hanwha systems/ICT, Seoul, Republic of South Korea.

Hi. I'm Inhyeok Yoo.

I have experience in various NLP project such as social media analysis, information retrieval, data analysis, text mining and sentiment analysis. You can check my C.V. here.

My goal is to be a GURU deep learning engineer in NLP area.

Tech Stack ๐Ÿ’ก

Conference ๐Ÿ“„

  • Yoo, I., Park, J. Y., & Kang, S.W. (2018), A Study of Correlation Analysis between Increase/Decrease Rate of Tweets Before and After Opening and a Box Office Gross. In proceedings of International Conference on Engineering and Science, ICENS, Tokyo, Japan 2018.07.31
  • Yoo, I., & Kang, S.W. (2018), Exploring correlation between social network service and box-office grosses from large-scale Twitter data. In proceedings of International Conference on Engineering and Science, ICENS, Sapporo, Japan 2018.01.31

PUBLISH ๐Ÿ“„

Awards ๐Ÿ†

  • ๋Œ€์ƒ, I-GPS ์ตœ์ข…์„ฑ๊ณผ ๋ฐœํ‘œํšŒ, ์ธํ•˜๋Œ€ํ•™๊ต ์‹ค์ „๋ฌธ์ œ ์—ฐ๊ตฌํŒ€.
  • Distinguished Paper Award, International Conference on Engineering and Science
  • 2018.07.30 ์žฅ๋ ค์ƒ, ๊ตํ†ต ๋น…๋ฐ์ดํ„ฐ ํ™œ์šฉ ์šฐ์ˆ˜๋…ผ๋ฌธ ๋ฐ ์•„์ด๋””์–ด ๊ณต๋ชจ์ „, ํ•œ๊ตญ๊ตํ†ต์—ฐ๊ตฌ์›
  • ๊ฒฝ์˜๋ถ€๋ฌธ์šฐ์ˆ˜ํ•™์ˆ ์ƒ, ๋Œ€ํ•œ์•ˆ์ „๊ฒฝ์˜๊ณผํ•™ํšŒ, ๋Œ€ํ•œ์•ˆ์ „๊ฒฝ์˜๊ณผํ•™ํšŒ์ง€
  • Distinguished Paper Award, International Conference on Engineering and Science
  • ์ตœ์šฐ์ˆ˜์ƒ, ์ œ 1 ํšŒ X-Corps ํŽ˜์Šคํ‹ฐ๋ฒŒ, ํ•œ๊ตญ์—ฐ๊ตฌ์žฌ๋‹จ ์ด์‚ฌ์žฅ
  • ์ตœ์šฐ์ˆ˜์ƒ, I-GPS ์ตœ์ข…์„ฑ๊ณผ ๋ฐœํ‘œํšŒ, ์ธํ•˜๋Œ€ํ•™๊ต ์‹ค์ „๋ฌธ์ œํ•ด๊ฒฐํŒ€
  • ๋Œ€์ƒ, I-GPS ์ค‘๊ฐ„์„ฑ๊ณผ ๋ฐœํ‘œํšŒ, ์ธํ•˜๋Œ€ํ•™๊ต ์‹ค์ „๋ฌธ์ œํ•ด๊ฒฐํŒ€

rag-tutorial's People

Contributors

inhyeokyoo avatar

Watchers

 avatar

rag-tutorial's Issues

AI hub ๋ฒ•๋ฅ /๊ทœ์ • (ํŒ๊ฒฐ์„œ, ์•ฝ๊ด€ ๋“ฑ) ํ…์ŠคํŠธ ๋ถ„์„ ๋ฐ์ดํ„ฐ EDA ์ˆ˜ํ–‰

Is your feature request related to a problem? Please describe.
๊ธฐ์กด์— ์‚ฌ์šฉ ์ค‘์ด๋˜ AI Hub ๊ธฐ์ˆ ๊ณผํ•™ ๋ฌธ์„œ ๊ธฐ๊ณ„๋…ํ•ด ๋ฐ์ดํ„ฐ์—๋Š” HTML ํƒœ๊ทธ ๋ฐ LATEX ๋“ฑ ๋…ธ์ด์ฆˆ๊ฐ€ ๋งŽ์•„ ๋‚œ์ด๋„๋ฅผ ์ข€ ๋” ๋‚ฎ์ถ”๊ธฐ ์œ„ํ•ด ๋ฒ•๋ฅ /๊ทœ์ • (ํŒ๊ฒฐ์„œ, ์•ฝ๊ด€ ๋“ฑ) ํ…์ŠคํŠธ ๋ถ„์„๋ฅผ ์‚ฌ์šฉํ† ๋ก ํ•จ.

๋”ฐ๋ผ์„œ ํ•ด๋‹น ๋ฌธ์„œ์— ๋Œ€ํ•œ ๋ถ„์„์ด ํ•„์š”ํ•จ.

Describe the solution you'd like
A clear and concise description of what you want to happen.

Describe alternatives you've considered
A clear and concise description of any alternative solutions or features you've considered.

Additional context
Add any other context or screenshots about the feature request here.

ํŒ€์› ์ง„ํ–‰์ƒํ™ฉ

์‚ฌ์šฉํ™˜๊ฒฝ Embedding Retrieval LLM
๋‚˜ (์ธํ˜) local SBERT (BM-K/KoSimCSE-Roberta) haystack Gemini-pro
์Šฌ๊ธฐ๋‹˜ colab BAAI/bge-small-en-v1.5
SBERT (distiluse-base-multilingual-cased-v1)
llama-index Replicate
(LLaMA, Mistral 7B, Solar)
์ง€๋กœ๋‹˜ colab - llama-index LLM-Camel-5b

์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ฆฌ์†Œ์Šค ๋ชฉ๋ก

Data

Note:

  • RAG ์—ฌ๋ถ€์— ๋”ฐ๋ฅธ ์„ฑ๋Šฅ ์—ฌ๋ถ€๊ฐ€ ์ž˜ ๋ณด์˜€์œผ๋ฉด ์ข‹๊ฒ ์Œ
    • RAG๊ฐ€ ์—†์„ ๊ฒฝ์šฐ hallucination ๋ฐœ์ƒํ•˜์—ฌ RAG์˜ ์„ฑ๋Šฅ์„ ๋ณผ ์ˆ˜ ์žˆ์–ด์•ผ ํ•จ.
    • ์—„๋ฐ€ํ•˜๊ณ  ์„ธ์„ธํ•œ ์ง€์‹์ด ์š”๊ตฌ๋˜๋Š” specificํ•œ ์ผ€์ด์Šค
  • ํ˜•ํƒœ๊ฐ€ ๋‹ค์–‘ํ•˜์—ฌ RAG ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ์—์„œ ์ œ๊ณตํ•˜๋Š” ๊ธฐ๋Šฅ์„ ์ถฉ๋ถ„ํžˆ ์‚ฌ์šฉํ•˜๊ณ  ์ฐจ์ด์ ์„ ํ™•์ธ
    • e.g., Table, list, hierarchy, chunking, prompt compression
  • PDFํŒŒ์ผ ๋“ฑ ์ „์ฒ˜๋ฆฌ์— ๋”ฐ๋ผ ์„ฑ๋Šฅ์ด ๋‹ฌ๋ผ์ง€๋Š” ๊ฒƒ์€ ์ œ์™ธ: ๊ด€๋ฆฌ ํฌ์ธํŠธ๊ฐ€ ์ฆ๊ฐ€ํ•จ
๋ฐ์ดํ„ฐ ์žฅ์  ๋‹จ์ 
๋„๋ฐฐ ํ•˜์ž ์งˆ์˜ ์‘๋‹ต ์ฒ˜๋ฆฌ : ํ•œ์†”๋ฐ์ฝ” ์‹œ์ฆŒ2 AI ๊ฒฝ์ง„๋Œ€ํšŒ
  • Hallucination์„ ์ผ์œผํ‚ฌ ์ˆ˜๋„ ์žˆ์–ด RAG ์œ ๋ฌด์— ๋”ฐ๋ฅธ ์„ฑ๋Šฅ์„ ๋ณด๊ธฐ ํŽธํ•จ (hard example)
  • MRC ๋ฐ์ดํ„ฐ๋ผ ์งˆ๋ฌธ๊ณผ ๊ทผ๊ฑฐ ๋“ฑ์ด ํ•จ๊ป˜ ์žˆ์Œ
์• ๋งคํ•จ
AI Hub ๊ธฐ์ˆ ๊ณผํ•™ ๋ฌธ์„œ ๊ธฐ๊ณ„๋…ํ•ด ๋ฐ์ดํ„ฐ
  • Hallucination์„ ์ผ์œผํ‚ฌ ์ˆ˜๋„ ์žˆ์–ด RAG ์œ ๋ฌด์— ๋”ฐ๋ฅธ ์„ฑ๋Šฅ์„ ๋ณด๊ธฐ ํŽธํ•จ (hard example)
  • MRC ๋ฐ์ดํ„ฐ๋ผ ์งˆ๋ฌธ๊ณผ ๊ทผ๊ฑฐ ๋“ฑ์ด ํ•จ๊ป˜ ์žˆ์Œ
๊ณ„์ธต๊ตฌ์กฐ, ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ ๋“ฑ์ด ๋ถ€์กฑ
๋ฒ•๋ฅ /๊ทœ์ • (ํŒ๊ฒฐ์„œ, ์•ฝ๊ด€ ๋“ฑ) ํ…์ŠคํŠธ ๋ถ„์„
  • Hallucination์„ ์ผ์œผํ‚ฌ ์ˆ˜๋„ ์žˆ์–ด RAG ์œ ๋ฌด์— ๋”ฐ๋ฅธ ์„ฑ๋Šฅ์„ ๋ณด๊ธฐ ํŽธํ•จ (hard example)
  • ๊ณ„์ธต ๊ตฌ์กฐ ๋“ฑ์ด ์žˆ์Œ
๊ทผ๊ฑฐ๋ฒ•๋ น์€ ์žˆ์œผ๋‚˜ ์งˆ๋ฌธ์€ ์—†์Œ. ์ ๋‹นํžˆ ์ƒ์„ฑํ•˜๋ฉด ๋ ์ˆ˜๋„์žˆ์„๋“ฏ

Embedding

Massive Text Embedding Benchmark (MTEB) Leaderboard๋„ ์ฐธ๊ณ ํ•ด๋ณผ ๊ฒƒ

Haystack์˜ SentenceTransformersTextEmbedder์—์„œ KorDPR ์‚ฌ์šฉํ•˜๊ธฐ

Is your feature request related to a problem? Please describe.

  1. KorDPR์˜ ๊ฒฝ์šฐ SKT์˜ KoBERT๋ฅผ ์‚ฌ์šฉ ์ค‘.
  • huggingface hub์—์„œ ํ•ด๋‹น ํŒŒ์ผ์€ ์‚ฌ๋ผ์กŒ์œผ๋ฏ€๋กœ ์ด๋ฅผ ํ•ด๊ฒฐํ•ด์•ผํ•จ
  • KorDPR์˜ ๊ฒฝ์šฐ BertModel์— pooler๋ฅผ ๋‹ฌ๊ณ  ์žˆ์œผ๋ฏ€๋กœ transformers์˜ DPRModel์„ ์‚ฌ์šฉํ•˜๋ ค๋ฉด ๋ณ€๊ฒฝ์ด ํ•„์š”ํ•จ
  1. haystack์—์„œ๋Š” SentenceTransformersTextEmbedder๋ฅผ ํ†ตํ•˜์—ฌ ์‚ฌ์šฉ์ด ๊ฐ€๋Šฅํ•จ
  • ์ธํ’‹๊ณผ ์•„์›ƒํ’‹์„ ํŒŒ์•…ํ•ด์•ผํ•จ
  • ๋ฌธ์„œ์—์„œ ๋””ํดํŠธ๋กœ ์‚ฌ์šฉํ•˜๋Š” sentence-transformers/all-mpnet-base-v2์˜ ๊ฒฝ์šฐ sentence transformer์˜ ๊ฒฝ์šฐ MPNetForMaskedLM๋ฅผ ์‚ฌ์šฉ ์ค‘
  • Nowadays, most of the models in the Massive Text Embedding Benchmark (MTEB) Leaderboard are compatible with Sentence Transformers. For example, if you use BAAI/bge-large-en-v1.5, you should prefix your query with the following instruction: โ€œRepresent this sentence for searching relevant passages:โ€

    • bge-large-en-v1.5์˜ ๊ฒฝ์šฐ BertModel ๋ฒ ์ด์Šค์ž„.

Describe the solution you'd like
A clear and concise description of what you want to happen.

Describe alternatives you've considered
A clear and concise description of any alternative solutions or features you've considered.

Additional context
Add any other context or screenshots about the feature request here.

TO-DO List:

TO-DO List:

์ตœ์ข… ๋ชฉํ‘œ ๊ด€๋ จ

RAG:

  • Retrieval๋กœ ์ฃผ๋Š” top-k ๋ฌธ์„œ์˜ ์ตœ์ ํ™”: LLM์ด ์ƒ์„ฑํ•  ๋•Œ ํ•„์š”ํ•œ ๋ฌธ์„œ์˜ ์ˆ˜๋ฅผ ์ตœ์ ํ™”
    • ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ™œ์šฉํ•  ์ˆ˜๋„ ์žˆ์„ ๊ฒƒ์œผ๋กœ ๋ณด์ž„.
  • RAG์—์„œ ์‚ฌ์šฉํ•˜๊ธฐ ์ข‹์€/์ ์ ˆํ•œ chunking ์‚ฌ์ด์ฆˆ๋ฅผ ์ฐพ๋Š” ๋ฐฉ๋ฒ• (AutoRAG)
  • ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ ํ™œ์šฉ (RAG Survey)
  • Multi-hop QA
  • Prompt compression
    • ์ œ๋ชฉ, ์†Œ์ œ๋ชฉ ๋“ฑ์„ ๋ฌธ์„œ ์•ž์—๋‹ค๊ฐ€ ๋ถ™์—ฌ์ฃผ๋Š” ์‹์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜๋Š” ์žˆ์Œ

Prompt:

  • Multi-turn ๋Œ€ํ™”
  • Prompting ์ž˜ ์ฃผ๋Š” ๋ฐฉ๋ฒ•

Optimization (Advanced)

  • Hybrid Search Exploration
    • lexical, semantic, vector search ๋“ฑ์„ ์กฐํ•ฉํ•˜์—ฌ ์ง„ํ–‰
  • Recursive Retrieval and Query Engine
    • Recursive retrieval involves acquiring smaller chunks during the initial retrieval phase to capture key semantic meanings. Subsequently, larger chunks containing more contextual information are provided to the LLM in later stages of the process.

  • StepBack-prompt
    • encourages the LLM to move away from specific instances and engage in reasoning around broader concepts and principles.

  • HyDE (Hypothetical Document Embeddings)
    • ์ƒ์„ฑ๋œ ๋‹ต๋ณ€์ด ์ฟผ๋ฆฌ(query embedding = document embedding)๋ณด๋‹ค embedding space์—์„œ ๋” ๊ฐ€๊น๋‹ค๋Š” ๊ฐ€์ •์œผ๋กœ ์ง„ํ–‰. ์งˆ๋ฌธ์— ๋Œ€ํ•ด hypothetical document๋ฅผ ์ƒ์„ฑํ•œ ์ดํ›„ ์ด์™€ ๋ฌธ์„œ ์‚ฌ์ด์˜ ๊ฑฐ๋ฆฌ๋ฅผ ์ธก์ •ํ•˜์—ฌ ํ™œ์šฉํ•˜๋Š” ๊ธฐ๋ฒ•

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    ๐Ÿ–– Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. ๐Ÿ“Š๐Ÿ“ˆ๐ŸŽ‰

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google โค๏ธ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.