AttentionX / onboarding-projects-team1

2 stars 0 forks source link

project 2 (Jun-Young Park) #8

Closed engineerA314 closed 1 year ago

engineerA314 commented 1 year ago

Abstract

haystack 을 이용한 Retrieval Augmented Generator + Chatbot

Transformer 논문과 RETRO 논문에 대한 QA가 가능한 chatbot 구현

Idea

질문이 들어오면 haystack 의 DocumentStore 와 EmbeddingRetreiver 를 이용해 질문과 가장 유사도가 높은 Chunk를 뽑고, 이를 chatbot의 prompt에 넣어줌

Prompt Design

Problems

  1. embedding dot product 기반 유사도가 적절하지 못함. 질문과 크게 관련없어보이는 정보들을 가져옴
    • 숫자가 들어있는 무의미한 chunk들이 선택되는 경우가 종종 있음.

image

  1. relative information chunks 는 유저에게 보이지 않는데, 이를 유저가 안다는 식으로 대답함. "in paragraph 1, ..." 와 같은 식으로 relative information chunks 내부의 내용을 인용하면서 대답하는 경우가 있음

  2. abstract 를 논문에서 주어진 것을 사용했는데 이거보다 더 좋은 방법이 있을 수도 있음

  3. 일상적인 대화가 불가능함

    Ask Question : Thank you. You helped me a lot.
    A :  I understand the information. I'm ready to answer the questions.

    위와 같은 식으로 prompt 에 의해 일상적인 대화를 할 수가 없음.

eubinecto commented 1 year ago

수고하셨습니당 ㅎㅎ