# Domain AI 應用：透過智能自動化改變行業

:crown: Speaker: <font color="#00bfff">**Leo Chu @Unition**</font>

{%hackmd 3_LMFJLvQV-rMOZRjKILnQ %}

:raised_hands: Slido: https://app.sli.do/event/62My7psHmtEaaKDXG8baBZ

###### ✦✦✦✦✦ Here we go ! ✦✦✦✦✦

## Movtivation
Traditional Customer Service - 客服
因客戶愈多, 所需要的客服愈多. 想提供線上AI客服

### Customer Service
Adv:
Problem-solving
Interpersonal interaction
Credibility - 有些客戶比較喜歡真人, 不會冷冰冰的

Dis-adv:
High Cost
Low Efficiency
Inconsistency
Limited Working Hours

--
### **AI** Customer Service
Adv:
Cost-effectiveness
Real-Time responses
Scalability
Automation

Dis-adv:
Lack of emotions and personalization
Limited responses
> 只能回應訓練的資識

---
## How to feed knowledge to GPT Model
1.Weight
2.Input

### Via model weights - fine-tuning
1. Like a lone term memory
2. High cost on prepare data
3. Fine-tuning is better suited to teaching specialized task or styles , and is less reliable for factual

### Via model inputs
1. message inputs are like short-term memory
2. open book exam

demo : ask with article

---
## Tokens

沒法一次給這麼多token , 就算是有也是挺花錢的.

What is token? : 處理文本的最小單位, 使用者輸入內容. e.g. 一個中文字2個token

Token Limit:
GPT3.5: 4,096 tokens
GPT4
GPT3.5-16k
GPT4-32K

## Embedding Data
1. Convert document/input to vector
做向量化時保持原本的特徵, 然後做比較.
2. Calculate similarity
我們怎麼做比較的呢? Consine Similarity

比對相似度的概念:
RBG model
Red : [255,0,0] , Pink [255,105, 180]
Consine = 0.77
> Consine做完的range 0 ~ 1, 0.77 算高的了. 意思是紅跟粉紅很像

---
## Embedding Search
1. Prepare search data
2. Search
3. Ask

### Prepare search data
1. Collect
2. Chunk
3. Embed
4. Store

### Search
1. Given a user question
2. Generate an embedding for the query
3. Rank the text sections
4. Insert the question and the most relevant sections
5. Return GPT's answer

Demo : Failed , Not working 來個掌聲.

在DB就2個欄 , content & 向量

---

# Prompt

需要精準的描述, 才能完成合理的任務

OpnenAI Prompt Types
1. System Role
2. Assistant Role
3. User Role

---

# Open AI

Why Azure Open AI

| |Embedding Search | Cognitive Search|
| -------- | -------- | -------- |
| Multi-lang | X | V |
| Cost for preprocessing Data | V | X |
| Support Differnt Type input | X | v |

Security

How Azure Open AI work

flow : Questoin -> Query from DB as `Knowledge` -> Prompt + `Knowledge`

Demo : Fail (請自行想像)

---

有提供API 做Embed 嗎?
yes , 且可以自己選擇等級 Ada or 達文西
可以處理一段文字.

Leo : 建議在演講前, 記得先在現場試過Demo先