ClassGPT

ChatGPT for my lecture slides

Built with Streamlit, powered by LlamaIndex and LangChain.

Uses the latest ChatGPT API from OpenAI.

Inspired by AthensGPT

App Demo

demo.mp4

How this works

1. Parses pdf with pypdf
2. Index Construction with LlamaIndex's GPTSimpleVectorIndex
   • the text-embedding-ada-002 model is used to create embeddings
   • see vector store index page to learn more
   • here's a sample index
3. indexes and files are stored on s3
4. Query the index
   • uses the latest ChatGPT model gpt-3.5-turbo

Usage

Configuration and secrets

1. configure aws (quickstart)

    aws configure

2. create s3 bucket named "classgpt"

3. rename [.env.local.example] to .env and add your openai credentials

Locally

1. create python env

    conda create -n classgpt python=3.9
    conda activate classgpt

2. install dependencies

    pip install -r requirements.txt

3. run streamlit app

    cd app/
    streamlit run app/01_❓_Ask.py

Docker

Alternative, you can use Docker

    docker compose up

Then open up a new tab and navigate to http://localhost:8501/

TODO

• fix ValueError: Could not parse LLM output issues
• Add ability to query on multiple files
  • Compose indices of multiple lectures and query on all of them
  • loop through all existing index, create the ones that haven't been created, and compose them together
  • references
    • Composability — LlamaIndex documentation
    • gpt_index/ComposableIndices.ipynb
    • Test Complex Queries over Multiple Documents
• Custom prompts and tweak settings
  • create a settings page for tweaking model parameters and provide custom prompts example
  • choose local or cloud storage version, so users don't have to setup AWS s3 and everything is downloaded locally
• deploy app on AWS

FAQ

Tokens

Tokens can be thought of as pieces of words. Before the API processes the prompts, the input is broken down into tokens. These tokens are not cut up exactly where the words start or end - tokens can include trailing spaces and even sub-words. Here are some helpful rules of thumb for understanding tokens in terms of lengths:

• 1 token ~= 4 chars in English
• 1 token ~= ¾ words
• 100 tokens ~= 75 words
• 1-2 sentence ~= 30 tokens
• 1 paragraph ~= 100 tokens
• 1,500 words ~= 2048 tokens

Try the OpenAI Tokenizer tool

Source

Embeddings

An embedding is a vector (list) of floating point numbers. The distance between two vectors measures their relatedness. Small distances suggest high relatedness and large distances suggest low relatedness.

For text-embedding-ada-002, cost is $0.0004 / 1k tokens or 3000 pages/dollar

• Embeddings - OpenAI API
• What Are Word and Sentence Embeddings?

Models

For gpt-3.5-turbo model (ChatGPTAPI) cost is $0.002 / 1K tokens

For text-davinci-003 model, cost is $0.02 / 1K tokens

• Chat completion - OpenAI API

References

Streamlit

• Increase upload limit of st.file_uploader
• st.cache_resource - Streamlit Docs
• Session State
• hayabhay/whisper-ui: Streamlit UI for OpenAI's Whisper

Deplyoment

• Streamlit Deployment Guide (wiki) - 🚀 Deployment - Streamlit
• How to Deploy a streamlit application to AWS? Part-3

LlamaIndex

• LlamaIndex Usage Pattern
• Saving index

Loading data

• PDF Loader
• llama-hub github repo
• document class
• PDFReader class

ChatGPT

• gpt_index/SimpleIndexDemo-ChatGPT.ipynb

Langchain

• gpt_index/LangchainDemo.ipynb
• OpenAIChat

Boto3

• boto3 file_upload does it check if file exists
• Boto 3: Resource vs Client
• Writing json to file in s3 bucket

Docker stuff

• amazon web services - What is the best way to pass AWS credentials to a Docker container?
• docker-compose up failing due to: error: can't find Rust compiler · Issue #572 · acheong08/ChatGPT
• linux - When installing Rust toolchain in Docker, Bash source command doesn't work
• software installation - How to install a package with apt without the "Do you want to continue [Y/n]?" prompt? - Ask Ubuntu
• How to use sudo inside a docker container?