AI Foundations - Chatbots & GPTs
How to Get Rich with AI Agents in 2025
名詞
- GPTs : Custom versions of ChatGPT
- Landing Page（著陸頁）的概念主要用於行銷和網路設計領域，指的是用戶透過某種方式（如點擊廣告、搜尋引擎結果、電子郵件鏈接等）進入網站後看到的第一個頁面。它的目的是吸引用戶注意，推動某種行動（例如購買、註冊或下載）。
- Funnel of GPTs 意指利用一系列 GPT 模型構建一個多層漏斗，每一層專注於特定任務或目標，逐步篩選、過濾或深化處理，直到達成最終結果。這種方法的靈感來自行銷中的「銷售漏斗」，但應用在 AI 系統的工作流中。
Fundation
The 2 ingredients of a GPT(AI agent)
- Prompting
- Knowledges
- Actions = Tools = Functions
How to build a good GPT
- Expert/Unique prompting
- Using private or hard-to-find data as custom knowledge
- Curating a unique mix of data
- Unique/Powerful action
Skill
Create GPTs for
- Freelance
- An agency selling to business
Skill List
Market research and use case identification
- Your experience
- Hijack experience
- Research for internet(Reddit, YouTube)
Data sourcing, Preparation and Curation
- Knowing how to find valuable data
- Prepare it for use in your agents for best performance
- Mix different data together
- Data type: Public, Hidden, Private
- Hidden Data from
Expensive or sought-after books
Data scraped from the web and combined in a unique way
Data collected from surveys, etc
Prompt Engineering
Tool Creation
- Free API
- No/Low Code Tool Builders
- Create/Host Yourself
GPT Marketing
- YouTube
- Landing Page
Creation
Reference
- GummySearch : Gummy Search 是一個專注於幫助用戶透過Reddit 社群進行市場調查和產品驗證的平台。
- RapidAPI : RapidAPI 是一個全面的API 市場，旨在連接開發者與各種API 供應商，讓使用者能夠輕鬆發現、連接和管理API。
- Relevance AI
How to Build No-Code AI Chatbots
introduce
New jobs
- Conversation AI designer
- AI & Chatbot developer
AI Automation Agencies(AAA) for
- Create powerful AI Chatbots and sell them to business
- Stack value by adding automations that integrate the chatbots into the business
- Package and sell to a specific niche
Types of Chatbots
- Old school ChatBots
- Been around forever
- Based on “if-then” decision trees
- Requires strict matches
- Binary choices > free form
- Pure AU Chatbots
- ChatGPT
- Free-form conversation
- run on an endless loop
- Complete opposite of old school
- Run on LLM APIs(e.g. ChatGPT API)
- Modern Chatbots
- Combines ‘Old School’ and AI
- Structured but with AI elements throughout
- Multi-purpose
- Can perform action(pull/push data etc)
- Can be extremely valuable
Two Types of pure AI Chatbots
- Prompted Assistants
- LLM API + Prompt
- Set into a mode or character
- Uses ‘Baked in’ knowledge of LLM
- Helpful but not valuable
- Custom Knowledge Chatbots
- LLM API + Prompt + Knowledge Base
- It can be very valuable
- Thouses of use case
- Not as simple as you think
Some Custom Knowledge Chatbots Example
- Basic Customer Support Chatbot
- Basic AI Persona Chatbot
- Basic Lead Generation Chatbot
- Basic Staff Training ChatBot
Software
Chatbot Software Categories(platform)
- Prototyping
- Complete Builders
- Tools(automation platform)
- Zapier
- stack AI
- make.com
Prototyping Software
- Good for quick demos(e.g. Loom - one handy tool)
- Proof of concept(POC)
- Limited customization
- Great way to get in the door with clients
Complete Builders
- Build anything you want!
- Basic to massively complex
- Perform actions(AIP calls, database, read/write)
- Easy development to major channels
Key Concepts
- Knowledge Bases
- Fine Tuning
- Prompting
- intent Classification(special chatbot, respond special thing)
- Chat History
- Deployment
Knowledge Base Basics
- A knowledge base is a database of test/number data chunks by similarity
- A chatbot can retrieve a small amount of interface most similar to the user query
- Only this info* is then passed to a model to help it generate an answer to their question(not ALL info)
- Fine Tuning
Issue
- Appearance VS Reality(In reality,this is NOT true)
With a correct setup knowledge base, you can create a chatbot that appears to know all of the information
Token Limit Pain(e.g: chatGPT has token and data source date limited)
- Model APIs have a fixed limit on information we can provide them with
- Impossible for them to look at all our data at once
- Need a retrieval(檢索) system to get the most relevant info to answer the user query
Data Chunking:Document –> Chunks–> Vector Database
Retriever Generator Model
- Chop data into small chunks
- Store chunks in vector database by similarity
- Receive user query
- Search the database for chunks similar to the query
- Return x number of chunks
- Combine chunks + prompt
- Send to LLM API
- Receive a response and send it to the user
Common Myths Debunked
- you are NOT training a model
- You are NOT fine-tuning a model
- You are NOT teaching the model anything
Fine Tuning
- Chat Models(What we typically use) can’t be fine-tuned yet
- The older models that could be retired
- Narrowing down a model to a particular use case
- Teaching it to identify a pattern and respond in a specific way
- E.g extracting alcohol expense from bank statement
- You are NOT adding new data to the model
Intent Classification
- Powerful tool for chatbot developers
- AI analyzes the use required to trigger actions
- Key component in modern chatbots
Development to where
- Webchat
- Message(Facebook)
- SMS
Building Chatbots LIVE
Agenda
- prototyping with Chatbase, Dante AI , Cody AI(1:12:48)
- Instruction to VoiceFlow(1:30:30)
- Customer Support Chatbot
- Staff Training Chat
- Lead Generation Chatbot
Customer Support Chatbot(1:38:00)
Whatsapp Staff Training(2:15:46)
+ VoiceFlow method for Whatsapp
+ make.com : for google sheet
Lead Generation(2:38:52)
+ VoiceFlow
+ make.com : for google sheet