RAG

Retrieval Augmented Generation
檢索增強生成

AI應用研究教育中心

真理大學

王柳鋐
2024/11

Last modified

2024/11/5

Outline

☞ Revolution of LLM

☞ Blooming of LLM

☞ More Than You Think, Lower Than Expected

☞ Practice using LlamaIndex & GPT

Revolution of LLM

Representation of Words

Training Paradigm

source1, source2, source3

word embedding

向量長度 =詞彙數量(e.g. length: 10000)

Network Architecture

輸入層

隱藏層

輸出層

Transformer, 2017/06

Word2Vec, 2013/01

Pre-trained, 2010/10

Hopfield, J. J. (1982). Neural networks and physical systems with emergent collective computational abilities. Proceedings of the national academy of sciences, 79(8), 2554-2558.

Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Computation, 9(8), 1735-1780.

Rumelhart, D., Hinton, G. & Williams, R. Learning representations by back-propagating errors. Nature 323, 533–536 (1986). https://doi.org/10.1038/323533a0

LeCun, Y., Boser, B., Denker, J. S., Henderson, D., Howard, R. E., Hubbard, W., & Jackel, L. D. (1989). Backpropagation applied to handwritten zip code recognition. Neural computation, 1(4), 541-551

CNN

LSTM

RNN

Novel Network Architecture: Transformer(1/3)

Sutskever, I., Vinyals, O., & Le, Q.V. (2014). Sequence to Sequence Learning with Neural Networks. ArXiv, abs/1409.3215.

圖片來源1

seq2seq model, 2014/09

Encoder-Decoder Model for NLP tasks

文字型任務

編碼器

解碼器

圖片來源

Novel Network Architecture: Transformer(2/3)

attention, 2014/09

Bahdanau, D., Cho, K., & Bengio, Y. (2014). Neural Machine Translation by Jointly Learning to Align and Translate. CoRR, abs/1409.0473.

seq2seq model, 2014/09

Attention: focus on specific parts of input while generating output

認知科學：選擇性注意力

圖片來源

Novel Network Architecture: Transformer(3/3)

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł. & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems (p./pp. 5998--6008), .

attention, 2014/09

seq2seq model, 2014/09

Transformer, 2017/06

Self-Attention: input interact with each other

上下文context

圖片來源, 圖片來源2

Transformer-based Language Models BERT(1/5)

Devlin, J., Chang, M., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. North American Chapter of the Association for Computational Linguistics.

attention, 2014/09

seq2seq model, 2014/09

Transformer, 2017/06

BERT, 2018/10

token

embedding

(Encoder-only)

(BERT base)

預訓練pre-training

prediction (probability)

下游任務downstream task

FNN

(微調, Fine-tuning)

dim: 768

Training Paradigm Shifting Pre-trained + fine-tuning

參考資料: Transfer Learning in NLP

遷移學習：預訓練(pre-trained) ＋目標領域資料集收集、訓練(fine-tune)

機器學習傳統做法

資料集1

資料集2

資料集3

任務1

任務2

任務3

兩階段: 非監督式、監督式

不同任務、不同資料集

Transfer Learning

來源領域任務

資料集S

目標領域任務

大量

資料集T

小量

預訓練unsupervised

微調supervised

Pan, S.J., & Yang, Q. (2010). A Survey on Transfer Learning. IEEE Transactions on Knowledge and Data Engineering, 22, 1345-1359.

pre-trained, 2010

Representation of Words one-hot encoding(1/3)

圖片來源1,圖片來源2

One-hot encoding

編碼：以向量表示

編號

詞彙

<SOS>

編號

0 0 0 0 0 0 0 0

<SOS> I played the piano

編碼

輸入層

隱藏層

輸出層

Word Embedding(詞嵌入)

字彙語意相近，編碼必須給予比較近的「距離」

Representation of Words word embedding(2/3)

Mikolov, T., Chen, K., Corrado, G.S., & Dean, J. (2013). Efficient Estimation of Word Representations in Vector Space. International Conference on Learning Representations.

Word2Vec, 2013/01

參考資料: Training Word2vec using gensim

Word Embedding

詞彙

維度

Word2Vec

Representation of Words word embedding: word2vec(3/3)

Word2Vec學習大量詞彙，將字詞對應到100-300維度的空間。

圖片來源, 圖片來源2

Transformer-based Language Models GPT-2 (2/5)

Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., & Sutskever, I. (2019). Language Models are Unsupervised Multitask Learners.

attention, 2014/09

seq2seq model, 2014/09

Transformer, 2017/06

GPT-2, 2019

BERT, 2018/10

decoder-only

117M Parameters

1,542M Parameters

圖片來源

Transformer-based LM Pre-trained + Fine-tune (3/5)

預訓練

微調

Transformer-based LM You Never Know What They Really Do (4/5)

Transform-based Model

Icon by Nikita Golubev

●●●

pre-trained data sets

NN1

Task 1

NN2

Task 2

NNn

Task N

target domain data sets

●●●

LLM

難以捕捉更高層次的語義概念
難以進行邏輯運算和因果推理
缺乏情境理解

Blooming of LLMs

Gemini Pro 1.5

GTP-4o

Claude 3.5 Sonnet

LLaMa 3.2

LLaMa 3.1 Sonar Large

Microsoft 365 Copilot

Open AI

GTP-o1 preview

https://paperswithcode.com/sota

LLM comparison Browse State-of-the-Art (1/3)

https://paperswithcode.com/sota

圖片來源

LLM comparison Browse State-of-the-Art (2/3)

Sept. 8th 2024

LLM comparison price, context, pros & cons (3/3)

圖片來源

LLMs are More Than You Think, but also Lower Than Expected

source1, source2, source3

Capabilities of Pattern Matching

Mirzadeh, I., Alizadeh-Vahid, K., Shahrokhi, H., Tuzel, O., Bengio, S., & Farajtabar, M. (2024). GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models.

One of the solutions: RAG

Hallucination

Prompt Engineering

Prompt Engineering Describing tasks via NL (1/3)

McCann, B., Keskar, N.S., Xiong, C., & Socher, R. (2018). The Natural Language Decathlon: Multitask Learning as Question Answering. ArXiv, abs/1806.08730.

attention, 2014/09

seq2seq model, 2014/09

Transformer, 2017/06

GPT-2, 2019

BERT, 2018/10

Prompt Engineering, 2018/06

A prompt is natural language text describing the task that an AI should perform.

It can be a query, specifying a style, providing relevant context, or assigning a role.

圖片來源

Prompt Engineering Describing tasks via NL (2/3)

attention, 2014/09

seq2seq model, 2014/09

Transformer, 2017/06

GPT-2, 2019

BERT, 2018/10

Brown, T.B., Mann, B., Ryder, N., et. al. (2020). Language Models are Few-Shot Learners. ArXiv, abs/2005.14165.

Prompt Engineering, 2018/06

Meta-learning: learning in context, not real training.

Few-Shot Learner, 2020/05

圖片來源

Prompt Engineering Describing tasks via NL (3/3)

attention, 2014/09

seq2seq model, 2014/09

Transformer, 2017/06

GPT-2, 2019

BERT, 2018/10

Few-Shot Learner, 2020/05

Wei, J., Wang, X., Schuurmans, D., Bosma, M., Chi, E.H., Xia, F., Le, Q., & Zhou, D. (2022). Chain of Thought Prompting Elicits Reasoning in Large Language Models. ArXiv, abs/2201.11903.

Prompt Engineering, 2018/06

Chain-of-Thought, 2022/01

Step-by-Step

Q+ \sum_{i} A_i

as Input

Retrieval Augmented Generation RAG (1/5)

attention, 2014/09

seq2seq model, 2014/09

Transformer, 2017/06

GPT-2, 2019

BERT, 2018/10

Prompt Engineering, 2018/06

Lewis, Patrick, et al. "Retrieval-augmented generation for knowledge-intensive nlp tasks." Advances in Neural Information Processing Systems 33 (2020): 9459-9474.

RAG, 2020/05

RAG for knowledge intensive tasks

1. parametric memory: a pre-trained seq2seq model

2. non-parametric memory: a dense vector index of Wikipedia

e.g. word embedding

RAG ReAct Reasoning and Action (2/6)

attention, 2014/09

seq2seq model, 2014/09

Transformer, 2017/06

GPT-2, 2019

BERT, 2018/10

Prompt Engineering, 2018/06

Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., & Cao, Y. (2022). ReAct: Synergizing Reasoning and Acting in Language Models. ArXiv, abs/2210.03629.

RAG, 2020/05

ReAct, 2022/10

Chain-of-Thought

Act 1: Search [Q]

Act 2: Search [Obs 1]

Act 3: Search [Obs 2]

Act 4: Finish [answer]

context

LangChain ReAct Design Pattern: High Level Abstraction (3/6)

attention, 2014/09

seq2seq model, 2014/09

Transformer, 2017/06

GPT-2, 2019

BERT, 2018/10

Prompt Engineering, 2018/06

RAG, 2020/05

ReAct, 2022/10

Setting, Resouces

Interactive UI for Work Flow Design

https://dify.ai/

source1, source2

LangChain + LangSmith

資料來源

圖片來源

LlamaIndex For Efficient Indexing & Retrieval (4/6)

https://www.llamaindex.ai/

圖片來源

LlamaIndex For Efficient Indexing & Retrieval

Source

RAG Research Framework (5/6)

RAG possible knowledge aware applications (6/6)

Ji, S., Pan, S., Cambria, E., Marttinen, P., & Yu, P.S. (2020). A Survey on Knowledge Graphs: Representation, Acquisition, and Applications. IEEE Transactions on Neural Networks and Learning Systems, 33, 494-514.

source

Zou, X. (2020). A Survey on Application of Knowledge Graph. Journal of Physics: Conference Series, 1487.

Legal AI

web crawler

Context
Query
Prompt

LLM

Output

Vector DB

source

RAG

關懷理論

Workflow icons created by Haca Studio - Flaticon

Scenario: 法律扶助

問責合規

來源透明

專家審查

defining an appropriate workflow

Practice RAG system from scratch

Context
Query
Prompt

LLM

Output

Vector DB

❶ Dataset

❸ Embedding

➍ Similarity

❷

➎ Reranking algorithm

➏

source

pip install llama-index

LlamaIndex Hello world (1/6)

1. 安裝Python套件

圖片來源1, 圖片來源2, 圖片來源4, 圖片來源5

●●●

2. 準備知識庫

支援各種檔案格式(參考官網文件)

3. 設定OpenAI API Key (可更換別的LLM)

LlamaIndex Hello world (2/6)

2. 準備知識庫

data資料夾：5個PDF檔

範例程式

知識庫

3. 設定OpenAI API Key (可更換別的LLM)

LlamaIndex Hello world (3/6)

3. 設定OpenAI API Key (可更換別的LLM)

LlamaIndex Hello world (3/6)

3. 設定OpenAI API Key (可更換別的LLM)

LlamaIndex Hello world (3/6)

輸入申請之API KEY

import os
from llama_index.core import (
    VectorStoreIndex,
    SimpleDirectoryReader,)

def setCurrentWD():
    abspath = os.path.abspath(__file__)
    dname = os.path.dirname(abspath)
    os.chdir(dname)

setCurrentWD()  # 設定工作目錄, 以免找不到data資料夾
# 1. Loading & Parsing
documents = SimpleDirectoryReader("data").load_data()

# 2. Indexing & vector store
index = VectorStoreIndex.from_documents(documents)

# 3. Query
query_engine = index.as_query_engine()
response = query_engine.query("Tell me about rag")
print(response)

LlamaIndex Hello world (4/6)

response = query_engine.query("Tell me about rag")

LlamaIndex output (5/6)

RAG models leverage a retriever to retrieve text documents based on an input query and use them as additional context when generating a target sequence. These models have been shown to achieve state-of-the-art results on various tasks such as open Natural Questions, WebQuestions, CuratedTrec, MS-MARCO, Jeopardy question generation, and FEVER fact verification. RAG models generate responses that are more factual, specific, and diverse compared to baseline models like BART. The retrieval mechanism in RAG plays a key role in improving results across different tasks.

實際回應

出處

LlamaIndex Cost (6/6)

$0.12

<$0.01

Embedding (indexes) 可以只算一次

Embedding預訓練是不是MultiLingual？效果差很多！

Hands on...

圖片來源

Retrieval Augmented Generation 檢索增強生成

Outline

Revolution of LLM

Novel Network Architecture: Transformer(1/3)

Novel Network Architecture: Transformer(2/3)

Novel Network Architecture: Transformer(3/3)

Transformer-based Language Models BERT(1/5)

Training Paradigm Shifting Pre-trained + fine-tuning

Representation of Words one-hot encoding(1/3)

Representation of Words word embedding(2/3)

Representation of Words word embedding: word2vec(3/3)

Transformer-based Language Models GPT-2 (2/5)

Transformer-based LM Pre-trained + Fine-tune (3/5)

Transformer-based LM You Never Know What They Really Do (4/5)

Transformer-based LM What happened to embedding? (5/5)

Transformer-based LM What happened to embedding? (5/5)

Transformer-based LM What happened to embedding? (5/5)

Blooming of LLMs

LLM comparison Browse State-of-the-Art (1/3)

LLM comparison Browse State-of-the-Art (2/3)

LLM comparison price, context, pros & cons (3/3)

LLMs are More Than You Think, but also Lower Than Expected

Prompt Engineering Describing tasks via NL (1/3)

Prompt Engineering Describing tasks via NL (2/3)

Prompt Engineering Describing tasks via NL (3/3)

Retrieval Augmented Generation RAG (1/5)

RAG ReAct Reasoning and Action (2/6)

LangChain ReAct Design Pattern: High Level Abstraction (3/6)

LlamaIndex For Efficient Indexing & Retrieval (4/6)

LlamaIndex For Efficient Indexing & Retrieval

RAG Research Framework (5/6)

RAG possible knowledge aware applications (6/6)

Legal AI

Practice RAG system from scratch

LlamaIndex Hello world (1/6)

LlamaIndex Hello world (2/6)

LlamaIndex Hello world (3/6)

LlamaIndex Hello world (3/6)

LlamaIndex Hello world (3/6)

LlamaIndex Hello world (4/6)

LlamaIndex output (5/6)

LlamaIndex Cost (6/6)

Hands on...

Retrieval Augmented Generation
檢索增強生成