TeleAssist - Web Server

Harness the power of Generative AI to reply your Telegram chats for you!

Introduction

Many companies and/or messaging platforms are already harnessing the power of AI to help users with replying messages. Just look at Gmail's nifty Smart Compose (email autocompletion) feature! They have even released a Smart Reply feature for Android Messages, and are slowly but surely improving their Machine Learning (ML) models to generate more appropriate response options to text messages received by users.

What if you are busy and don't have time to even choose messages to send to those NPC characters in your life? What if instead of simply suggesting responses, the model simply picks the best one and sends it directly? What if we are bored of the mundane and want to live a little more chaotically?

Features

• Bring Your Own Model
  • currently only HuggingFace models are supported right out of the box, though you can easily amend the code of the model loading class to fit your desired model
• Message Batching
  • To account for people's tendency to send multiple messages at once, messages within 9 seconds of one another from a single peer will be batched into one "question" to be answered by the model
• Option to add an introductory and/or farewell message
  • for responsible users like yourself to inform your peers that it is in fact a model that is speaking to them
  • able to specify a reason why you are busy, which will be included in the introductory message
• Only responds to trusted users and groups
  • Pre-configured to only reply to people in your contacts list, or ...
  • group chats (when you are tagged) where everyone in the group is in your contacts list
  • Will not to bots or unknown spam users
  • Being tagged in channels will NOT trigger this app's functionality
• Paired with an intuitive frontend interface
  • TeleAssist - Telegram Bot is a tool that silences your notifications and tells your model to take over replying while you are busy working on your tasks!

Video Demo

(to be updated!)

Limitations

• Only supports text mediums
  • Photo and sticker support are planned
• Not a fully-fledged chatbot
  • Does not "remember" past conversation turns ie. only replies directly to peer's current batch of messages

Requirements

Note that the functionality of this TeleAssist - Web Server is meant to be paired with the TeleAssist - Telegram Bot. It acts as the front-end user interface to the functionality provided by this app.

• Python 3
  • Recommended: >= 3.10.4
• Host Machine
  • RAM: >= 300MB to ensure optimal performance (does not include usage by model)
• MongoDB (optional)
  • the app does make a few helper function calls to log information into a local mongodb instance, but you could easily refactor and remove these helper functions from the codebase if you do not wish to log your chats
• Docker (optional)
  • Dockerfile provided if you wish to containerize it

Setup

1. Obtain Telegram api_id and api_hash

   • Link to instructions on Telegram docs

2. Clone repository

   > git clone https://github.com/lihaan/TeleAssist-Web.git

   > cd TeleAssistWeb/web_app

3. Create message_templates.py in ./api directory

   • This python file will contain the messages that you wish the app automatically sends at the start / end of the conversation
   • Refer to samples_messages_templates.txt for examples

4. Supply model, tokenizer, generation configuration into the ./text_generation_models directory

   • Example folder and naming structure:

     - text_generation_models/
         - distilgpt2_model_v1/
             - tokenizer.json,
             - vocab.json
             - special_token_map.json
             - etc
         - distilgpt2_tokenizer_v1/
             - config.json
             - pytorch_model.bin
             - training_args.bin
             - etc
         - contrastive_search_v1/
             - generation_config.json
     

5. Create a .env file with the following parameters

   • API_ID (Telegram api_id)
   • API_HASH (Telegram api_hash)
   • MY_PHONE (Phone number used for Telegram, in international format)
   • MY_HANDLE (telegram handle eg. @noobmaster69)
   • TOKENIZER_DIR (eg. distilgpt2_tokenizer_v1)
   • MODEL_DIR (eg. distilgpt2_model_v1)
   • GENERATION_CONFIG_DIR (eg. contrastive_search_v1)

6. Install dependencies

   • Recommended to create a virtual environment (eg. venv) before installing
     • requirements.txt is found at ./requirements.txt

       > pip install -r requirements.txt

     • Install PyTorch separately
       • Refer to PyTorch Get Started Locally for the appropriate way to install torch based on your OS

Run

1. Launch MongoDB instance
   • Either do this, or remove in code all function calls imported from mongoDB.py
   • Default URL: mongodb://localhost:27017
2. Run locally

   > python web_app/api/__init__.py

Build and Deploy (with Docker)

Web app

• Dockerfile is provided for reference at /web_app/Dockerfile

• If you are using TeleAssist - Telegram Bot, or opting to retain the functionality of the MongoDB logging, you should consider placing all docker containers in the same network

• Final list of commands to build and run:

  > docker network create tele-assist-network

  > docker build -t tele-assist-web .

  > docker run --name web --net tele-assist-network -p 5000:5000 -d tele-assist-web

MongoDB server

• Take note to amend the URL of the MongoDB database in web_app/api/mongoDB.py from localhost to [name_assigned_to_mongoDB_container]. In the command below, the name assigned to the docker container is db.
• Attach a docker volume to enable data persistance within container
  • Automatically created when a container is started with a volume that does not yet exist
• Command to build and run: (assuming network has been already created)

  > docker run --name db -p 27017:27017 --network tele-assist-network --mount source=tele-assist-volume,target=/data/db -d mongo:6.0.6

Server Maintenance

(to be updated!)

Contributing

Although TeleAssist is simply a pet project of mine, I am super keen on improving it and expanding its capabilities. Feel free to raise issues to provide feedback if you have any. If this project has helped you in any way, do give it a star or buy this struggling university student a coffee! Thank you!

~ Li Han