INDEX

###########################################################
2023-12-20 07:45
###########################################################

How to build chatbots
https://www.youtube.com/watch?v=jCoH82LPgdk

AAA (AI Automation agencies) - create packages of AI to sell to niches
Custom Knowledge Chatbots are the items of value - e.g. staff training
Prototyping: Chatbase, Dante, Cody - quick demos, POCs, limited customise
Builders: Botpress, Voiceflow - flowchart nocode bots - can call APIs
Tools: Zapier, Stack, Make - automate direct actions

Chatbot can retrieve small amount of data from knowledgebase at a time
Limited message sizes (token limits) - need to get only the most relevant data
Chunking = break document into pieces - store in vector db
Vector db stores similar data close together - have regions of similarity

Give the chatbot a context (role) and then a prompt (the content)
	import openai
	openai.ChatCompletion.create(
	  models='gpt-3.5-turbo'
	  messages=[
	    {"role": "assistant", "content": "What is a language model?"),
	    {"role": "user", "content": "What is chatgpt?")
	  ]
	)

Prompted AI query: Knowledge chunks + User history + User query
Intent Classifier: categorise queries e.g. ask for reviews or orders or data
Can deploy to webchat, whatsapp, messenger, SMS etc.
Agents have lists of tasks, use actions, evaluate and respond to users

Chatbase: New Chatbot -> Add data (has 400k character limit)
Gives a basic prototype based on a document or website
Settings -> OpenAI -> Base Prompt (initial prompt to always use)
Integrations -> Add to services or Embed on Site (in an iframe)

Dante: New knowledge base -> Add URLS/files - similar to Chatbase
Has more features like initial message, colours, styling changes etc.

Cody: Lets you manage knowledge in folders - link bots to folders
Content -> FOLDER -> Add data -> Bots -> New bot -> ... -> Add knowledge
Has pre-prompt presets and ratio sliders for content:history:response

Voiceflow is the main program - $50/month - can use templates too
New Assistant -> Build AI Assistant -> Web Chat -> Create
Knowledgebase file should be the most dense q&a you can make
Use chatbot Claude V1 - use chatgpt to write your initial prompt for you
Essentially build a flowchart of message-listen-prompt 
	Answer the user's question: {last_utterance}

Can build loops - prompt then answer, without complexity
Choice blocks capture intents triggered by key phrases ("utterances")
Can use intents like functions - e.g. go to "review" flow
Buttons give users specific answers to give
When using intents, go to "training" and retrain it - optimises a bit
Need to make sure within intents you tell it to only scope locally
Can get it to run javascript and store response in variables/flows