Skip to content

Latest commit

 

History

History
60 lines (38 loc) · 2.91 KB

readme.md

File metadata and controls

60 lines (38 loc) · 2.91 KB
Raw

Impersonator

Impersonator lets you chat with an AI simulation of any author, blogger or person as easily as copy-pasting texts they have written into a folder!

Installation

You can run the following pip command to install the dependencies (I recommend using a conda or venv environment to isolate the dependencies):

python3 -m pip install unstructured[local-inference] openai tiktoken bs4 faiss-cpu langchain

The nltk package (a subdependency) will require the following models:

python3 -m nltk.downloader punkt averaged_perceptron_tagger

Finally, you will need to install Detectron2 if you want to process inputs with pictures. This can be tricky but, you can skip this step by avoiding any pictures in your input files.

You can now git clone this repository wherever you want and start to use it!

Usage

Starting

To start the program, put your OpenAI API key in the environment (you can also type it at the beginning of the program) then start impersonator.py.

It will display a list of personas available (see this page for ways to add personas). If you select a persona that has never been used, it will first generate it from the data (which will take time proportional to the amount of data in the texts_by/texts_about subfolders of your persona, probably less than five minutes).

Once your persona is loaded, you can start chatting with it!

Special commands

You have a handful of special commands available. To use them, type one of those commands in the chat instead of your message:

  • FREE lets the persona extrapolate information, filling holes in its knowledge, from now on (the default)
  • STRICT avoids extrapolation from now on
  • CHECK will run a fact checker on the persona's latest affirmation,
  • SOURCE will display the text extracts used to synthesize the answer,
  • REDO will regenerate the latest answer,
  • EXIT will terminate the chat.

Documentation

Potential improvements

  • use the chatGPT API to reduce costs and simplify the code

  • have the persona be automatically regenerated if one touches the files in its folder (at best it should be updated rather than recomputed)

  • add the possibility of having multiple personas interacting in a single chat

  • add a persona-hub and the possibility to download from the hub

  • Add a non-shell UI

Do not hesitate to submit pull requests to this repository if you find improvements or good alternative prompts!