CescoGPT: a multipurpose CLI and Telegram bot, powered by OpenAI API

Overview

This application, written in Rust, enables users to utilize ChatGPT for a variety of purposes:

1. for plain, standard queries;
2. to practice conversation in a foreign language;
3. to correct and improve texts;
4. to summarize texts, choosing the output language and proficiency level;
5. to translate SRT movie subtitles into any desired language;
6. to transcribe audio files into text;

Configuration

Setting the API key

In order to use the bot you need a paid openAI API key, which must be set in the environment variable OPENAI_API_KEY:

export OPENAI_API_KEY="sk-__YOUR_API_KEY_HERE___"

Setting up the assistants

To get the bot up and running, you'll have to create five assistants and assign them the following names and instructions.

Generic ChatGPT

Let's chat.

Language Practice

You are CescoGPT, an AI to practice conversation in foreign
languages. You always reply in the current foreign language,
by 1. producing the correction to the previous message that you
received within <correct_me> and </correct_me> delimiters, formatting
it in this way: {Word for \"Correction\" in the foreign language}:
{corrected message}, 2. replying to the message and 3. you always end
your response with a related question.

Correct Text

You are CescoGPT, an AI to correct and improve texts.  You always
reply by producing the correction to the previous message that you
received within <correct_me> and </correct_me> delimiters, formatting
it without using the delimiters.

Summarize Text

You are CescoGPT, an AI designed to summarize texts. You always reply
by providing a summary of the original text that you receive within
<summarize_me> and </summarize_me> delimiters, formatting it without
using the delimiters. All the input texts you receive refer to the
same article, so remember them when you receive and summarize new
pieces of text.

Translate Subtitles

You are CescoGPT, an AI to accurately translate movie subtitles
between different languages. The subtitles are given as values of a
dictionary, codified as a JSON. You copy the keys of the dictionary
verbatim, while translating the values. You try to translate each
single string of the json on its own. However, if you merge
consecutive entries, you must preserve the first key of the merged
entries while dropping all the others. You translate all the text you
are given as input, without omitting any part. If an input sentence is
empty or too short for translation, you just reproduce it verbatim,
without changes. Your output is in JSON format, like the input.

Here's an example, translating into Italian. Input:
```
{"000obxmO": "Hi, how are", "001Lfyqd": "you?", "002aC3nE": "Fine, thanks."}
```
Desired output:
```
{"000obxmO": "Ciao, come va?", "002aC3nE": "Bene, grazie."}
```

Note: To reduce the risk of misformatting when translating subtitles, it is advisable to significantly lower this assistant's temperature, for example, from 1 to 0.01.

Filtering the user access

The configuration file conf/defaults.toml must contain an id_whitelist field, which can be filled with a list of allowed Telegram user_ids:

id_whitelist = [
  123456789,  # myself
  987654321,  # my cat
]

Note: If the list is left empty, no filtering will be applied (i.e., all users will be able to use the bot). Unless you feel particularly generous, it is highly recommended to populate the list with the authorized users and to set appropriate spending limits for your OpenAI API keys.

Installing the binaries

Assuming you have cargo correctly set up, to install all the binaries simply run:

cargo install --path .

Running the bot

The Telegram bot can be run with;

TELOXIDE_TOKEN=123_YOUR_TELEGRAM_BOT_TOKEN_567 \
cesco-gpt-bot

To enable logging of debug information, also set the RUST_LOG=debug environment variable.

Running the CLI

A command line interface is also available, to access the API straight from the shell. Just run with:

cesco-gpt generic  # generic ChatGPT prompt
cesco-gpt language-practice german b2  # practice B2 German
cesco-gpt correct --native  # correct and rephrase as a native speaker
cesco-gpt summarize italian c2  # summarize a text into C2 Italian
cesco-gpt -h  # get detailed help

The CLI concatenates consecutive lines and sends them as a message once an empty line is encountered (i.e., press enter twice to send the message). An empty message ends the conversation.

Logging the conversation

To log your CLI conversation to a text file, you can simply use the tee command, as done in this example:

cesco-gpt correct | tee /tmp/gpt-log.txt

Image generation via DALL-E 3

A CLI program for generating images from text prompts using DALL-E 3 is also available. To view the syntax, use the following command:

dalle-create -h

Subtitle tranlation

To translate movie subtitles (in SubRip SRT format), run the associated CLI program, for example:

translate-subs /tmp/original.deu.srt /tmp/translated.eng.srt english

This program can also make use of parallelism to improve the speed of computation. For details, run:

translate-subs -h

Speech to text

To transcribe audio files, run the associated CLI program, for example:

speech-to-text /tmp/audio.m4a /tmp/script.txt

The program also supports custom prompts and output in SubRip SRT format. For details, run:

speech-to-text -h

Language customization

If you want to modify the default available languages, just edit the Lang enum which is found in src/talks/lang_practice.rs.

Known problems

• Sometimes the OpenAI API may hang and only display three dots or an end-of-text symbol ␃ as a response. If this happens, you can try copying and pasting again your latest message and continue the conversation. If that doesn't work, you can use the command /restart to restart the Telegram conversation.
• Messages longer than 4096 characters are automatically split in shorter ones by Telegram, and treated as independent messages by this bot.
• Sometimes when summarizing a text ChatGPT may forget which output language it was supposed to use and switch to either the input one or English.

Author

CescoGPT is developed by

• Francesco Versaci francesco.versaci@gmail.com

License

CescoGPT is licensed under the under the Apache License, Version 2.0. See LICENSE for further details.