Dec 2023 Generative AI Intro Presentation

In this presentation we'll create some poems about Star Wars characters. For the final act, we'll build a vector database from a Wookieepedia site dump, and then query it with RAG.

Prereqs

• AWS account with Bedrock and some models enabled
• AWS SSO (if you don't have this, use regular old AWS CLI auth. See the docs for help.)
• node 20 / npm
• 7zip

Before the show

1. npm install
2. Create a dotenv file: touch .env && echo AWS_PROFILE={your AWS profile name} >>.env && echo AWS_DEFAULT_REGION={your AWS region} >>.env
3. Log in to your preferred AWS account: aws sso login
4. Create the Wookieepedia vector database:
   1. mkdir data
   2. download into /data from https://starwars.fandom.com/wiki/Special:Statistics (database dumps)
   3. Unzip with 7zip
   4. Extract with https://github.com/attardi/wikiextractor
   5. node act3-preload.mjs You'll need to comment out the kill switch first. This takes several hours, and you'll spend about $6 on Bedrock embeddings. You could also run Ollama or similar locally instead of using Bedrock.

There are usage examples at the top of each act*.mjs script. The /refs dir contains reference copies of several of the act*.mjs scripts with extra stuff commented out, so you can easily copy/paste it in during your demo.

Run of show

Act 1 - Basic LangChain

1. node act1.mjs
2. Change Yoda to Darth Vader
3. Change haiku to sonnet -> see maxTokens gotcha! (Then add maxTokens prop.)
4. Change temperature
5. Change model (copy/paste from refs/act1.mjs)
6. Demo debugging with LANGCHAIN_VERBOSE=true env var

Act 2 - Prompts

1. node act2.mjs "Yoda"
2. Play around with different voices
3. Demo LangChain step debugging (copy/paste from refs/act2.mjs)
4. Show alternate output parser: node act2-1.mjs "Yoda"
5. Uncomment parser instructions to see what gets passed to the prompt

Act 3 - RAG

1. node act3.mjs "What color is Yoda?"
2. Change model
3. Show the more detailed version of the chain (copy/paste from refs/act3.mjs)
4. Final question: "Who shot first?"