# 0613 AI contest
## prompt plugin API

#### point
- 明確指示模型要怎麼做
- 用分隔符號
- 明確的step,長度
- RAG Pattern⇒透過檢索的方式
- 把複雜的Task簡單化
- 先判斷意圖
- 再嘗試解決問題
- 在system message分類化的做判斷(可以有範例)
#### Entity Extraction > Summarization > Sentiment Analysis
#### Clearly statment
- like: give me som summary in 5 viewpoints
#### Can use template
- Set additional command in template for implement a user interface, but user have no idea about it
#### Few shot
- No need to pretrain
- Use it with few data and content
#### Split ChatGPT task into subtask
- Can ask it and ask to summerize the previous conversation
- Ask it to remove previous content, and start another question
### Fine tune
#### Use it for better and complex classifaction
- Not first choice (try prompt engineering firstly)
- Slove complex solution:
- Can upload a solution and ask it to compare it.
- Don't tell me the solution is true on not
- After comparing, ask it to evaluate the quilty
- How many 點(point) is match

### System message

### ChatGPT limitation
- no fine tuning, but Davinchi can do
- computing resource limitation (for chatting type)
- speak nonsense, but user can restrict the boundary

### Azure OpenAI Service

### Azure AI studio
#### ONLY for developer, not normal user

- built-in search for data pre-summarize(choose better result)
- built-in keyword searching
- built-in semantic searching
- built-in "@@@" searching
- Data resource
- Azure Machine Learning Prompt Flow
- Azure AI Content Safety: Check if the content is safe
- Deploy
#### Content Safty
- same as Azure AI API
- limitation
- rule violation

## Q&A
[1] 語意式的訓練，需要切斷嗎? 切斷的話有甚麼優缺點? -> ChatGPT的token = 8192，做embadding應該切開來，做 chunk，基本上沒上限(非硬性限制)
[2]