Slide 1

Slide 1 text

Bilge Yücel TJC & DefineX Mastering Generative AI From Word Embeddings to Custom AI Agents

Slide 2

Slide 2 text

01 - What is it? ● 🥑 Developer Relations Engineer at deepset ● 🏗 Open source LLM Framework: Haystack ● 📍 Istanbul, Turkey Bilge Yücel Twitter: @bilgeycl Linkedin: Bilge Yucel GitHub: @bilgeyucel Hi! 👋

Slide 3

Slide 3 text

NLP?

Slide 4

Slide 4 text

NLP?

Slide 5

Slide 5 text

NLP?

Slide 6

Slide 6 text

NLP - Natural Language Processing

Slide 7

Slide 7 text

Agenda 01 - Text Embeddings 02 - Vector Databases 03 - Retrieval 04 - LLMs 05 - RAG 06 - Agents & Function Calling

Slide 8

Slide 8 text

01 Text Embeddings

Slide 9

Slide 9 text

Text Embeddings/Text Vectors to be or not to be ● Manageable by computers ● Different techniques: ○ Sparse: TF-IDF, BM25... ○ Dense: Trained models (Sentence Transformers, Cohere, OpenAI...) ● Come in different dimensions

Slide 10

Slide 10 text

Text Embeddings - 2 dimensional

Slide 11

Slide 11 text

Text Embeddings - Semantic Similarity

Slide 12

Slide 12 text

02 Vector Databases

Slide 13

Slide 13 text

● Databases that store high-dimensional vectors ● Optimized for vectors: ○ Vector search ○ CRUD operations ○ Metadata filtering Vector Databases

Slide 14

Slide 14 text

Databases https://docs.haystack.deepset.ai/docs/choosing-a-document-store

Slide 15

Slide 15 text

03 Retrieval

Slide 16

Slide 16 text

Retrieval Query ● Getting the most relevant information to the query ● Used for semantic search, question answering and more

Slide 17

Slide 17 text

04 LLMs

Slide 18

Slide 18 text

Large Language Models (LLMs)

Slide 19

Slide 19 text

Large Language Models (LLMs) ● Big language models ● More data & more parameter ● Prompt → Human-like output ● Text generation: summarization, generative QA, writing code, chat…

Slide 20

Slide 20 text

Model Providers - Which one?

Slide 21

Slide 21 text

LLMs being LLMs ● Not conscious ● Use carefully https://x.com/hrrcnes/status/1793335665117044990

Slide 22

Slide 22 text

LLMs know everything?

Slide 23

Slide 23 text

LLM: Limitations ● LLMs do not know the answer to everything ● But they are good at following instructions ● We can help them in their task by giving them the relevant context + instruction

Slide 24

Slide 24 text

05 Retrieval Augmented Generation

Slide 25

Slide 25 text

Retrieval Augmented Generation

Slide 26

Slide 26 text

Retrieval Augmented Generation

Slide 27

Slide 27 text

● Haystack is an open source Python framework for building production-ready LLM applications ● Prototyping, evaluation, deployment, monitoring… ● Building blocks: Components & Pipelines Haystack

Slide 28

Slide 28 text

Haystack

Slide 29

Slide 29 text

Indexing Pipeline

Slide 30

Slide 30 text

Indexing Pipeline

Slide 31

Slide 31 text

Indexing Pipeline

Slide 32

Slide 32 text

Indexing Pipeline

Slide 33

Slide 33 text

RAG Pipeline

Slide 34

Slide 34 text

RAG Pipeline

Slide 35

Slide 35 text

RAG Pipeline

Slide 36

Slide 36 text

RAG Pipeline

Slide 37

Slide 37 text

RAG Pipeline

Slide 38

Slide 38 text

RAG Pipeline

Slide 39

Slide 39 text

No content

Slide 40

Slide 40 text

06 Agents

Slide 41

Slide 41 text

Agents ● Complex systems (multiple LLMs) ● Prompting: Chain of Thought/ReAct ● Generate accurate responses to complex queries ● Might use tools for answers ● Use cases: chat bot, personal assistants

Slide 42

Slide 42 text

Agents - CoT Chain-of-Thought Prompting Elicits Reasoning in Large Language Models - https://arxiv.org/pdf/2201.11903

Slide 43

Slide 43 text

Agents - CoT Chain-of-Thought Prompting Elicits Reasoning in Large Language Models - https://arxiv.org/pdf/2201.11903

Slide 44

Slide 44 text

REACT: SYNERGIZING REASONING AND ACTING IN LANGUAGE MODELS - https://arxiv.org/pdf/2210.03629 Agents - CoT

Slide 45

Slide 45 text

Agents - Reason + Act (ReAct) REACT: SYNERGIZING REASONING AND ACTING IN LANGUAGE MODELS - https://arxiv.org/pdf/2210.03629

Slide 46

Slide 46 text

Agents - ReAct Prompt

Slide 47

Slide 47 text

06.5 Function Calling

Slide 48

Slide 48 text

Function Calling ● Passing prompt + functions/tools (name, description, parameters) ● User query -> function name + JSON object for arguments ● Use arguments to call the function ● Voila! 🎉

Slide 49

Slide 49 text

Function Calling - Tools https://arxiv.org/pdf/2205.00445

Slide 50

Slide 50 text

Function Calling - Tool definition

Slide 51

Slide 51 text

Function Calling

Slide 52

Slide 52 text

Agents Search Google and answer: why did Elon Musk sue OpenAI?

Slide 53

Slide 53 text

Agents Search Google and answer: why did Elon Musk sue OpenAI?

Slide 54

Slide 54 text

Agents Search Google and answer: why did Elon Musk sue OpenAI?

Slide 55

Slide 55 text

Agents Search Google and answer: why did Elon Musk sue OpenAI? Elon Musk sued OpenAI for breaching its founding agreement and diverging from its original nonprofit mission….

Slide 56

Slide 56 text

Agents Using TinyURL create a shortened URL from this long URL https://haystack.deep set.ai/blog/haystack- 2-release https://tinyurl.com/hjr6yttr

Slide 57

Slide 57 text

Agents

Slide 58

Slide 58 text

Recap 01 - Text Embeddings 02 - Vector Databases 03 - Retrieval 04 - LLMs 05 - RAG 06 - Agents & Function Calling

Slide 59

Slide 59 text

@bilgeycl Bilge Yücel bilgeyucel Thank you! haystack.deepset.ai