colemanja91 / chatgpt-roleplay

README

This app is a helper for using ChatGPT to create entertaining, long-lived role-play characters.

I do not intend to run this app in a "production" (live, on-web) environment, due to OpenAI costs. But, I hope folks take the opportunity to run this locally.

The front-end is built in React: https://github.com/colemanja91/chatgpt-roleplay-ui

What it does

• Uses chat completion "system" messages to set the rules that the AI must follow
• Sends input prompts to OpenAI, along with recent message history and system message
  • Recent history is cutoff after some point based on the OpenAI max token limit
• Stores message history and system message in database for future reference
  • Allows for long-lived characters
• Enables variable temperature
  • Temperature controls how deterministic or random the output should be
  • Default of 1 is used (same as ChatGPT)
  • Enabling the appropriate option on message send will randomly select a valid temperature value to use
  • Can help make things more interesting over time by ensuring ChatGPT doesn't always send easily-guessable responses
• Serves functionality as a GraphQL API

Example 1 - System Prompt + Most Recent Messages

• Your character has the following state:
  • System message of 500 tokens
  • Message history of 100 messages (both user and assistant)
    • Each message length is 50 tokens (simplification for the sake of example, in reality this will be variable)
  • New input message of 150 tokens
• In order to meet OpenAI's API token size limits, the following will be sent:
  • The system message
  • The new input message
  • The most recent 32 messages from history
  • Total request token size: 3,850

Setup

Ruby Version

.ruby-version defines the current version; rbenv is recommended for version management.

Postgres

This app was developed against PostgreSQL 14.

Redis

Doesn't really matter what, just get it running.

Local setup

Install dependencies:

bundle

Setup databases:

bundle exec rails db:create
bundle exec rails db:migrate

Running tests

bundle exec rspec

Credentials

In the OpenAI developer console, create a new access token and set it in your environment as OPENAI_ACCESS_TOKEN. If you want to use a different environment variable, you can update the setting value in config/settings.default.yml.

Running App

bundle exec rails s

GraphQL queries can be routed to http://127.0.0.1:3000/graphql.

Graphiql is mounted at http://127.0.0.1:3000/graphiql and can be used for schema introspection and testing.

TODO

Initial development is happening during Twitch streams, and we'll use this as our backend work tracker for now.

Now

Next

• Generate a summary message from a range of message history
  • Ask ChatGPT to summarize history in 1 - 2 sentences and provide an importance score of 0 - 10
  • Store
    • Raw response (for debugging)
    • Response without importance score
    • Importance score
• Add Sidekiq + Redis for background processing
• Enqueue a job to generate summary messages after every message send
  • Should be idempotent if no summary is needed
• Include summaries in system message
  • Only summaries generated from messages before current message range
  • Only the top 3 summaries ranked by ChatGPT-determined importance and recency
  • Toggleable via API

Later

• Explore using GraphQL Subscriptions for messages
  • If it works, move OpenAI calls to a background job
• Dockerize?