PrivateGPTBot

Smart Telegram Bot based on ChatGPT

An asynchronous bot with multiple characters for accessing the gpt-3.5 and gpt-4 API (chatgpt) via your OpenAI key. It can work with documents and the internet, and interprets images using BLIP.

Bot's address

t.me/personalized_ai_bot - you can try it right now if the bot is in global mode.

Features

• Full Asynchronous: With the help of aiogram, the bot is fully asynchronous and actively uses multithreading.
• Function-calling support: Working with any number of documents, as well as website scraping.
• Multilingual interface: Automatic adaptation of the interface to the user's language.
• Adaptive truncation of history: The bot is capable of adaptively processing and truncating the history of messages according to the model's context size.
• Image processing: The bot can perceive and process images.
• Message customization: The bot provides the opportunity for full customization of all messages.
• Character creation: Ability to create a unique character within a session.
• System of tokens packages: Ability to control tokens consumption via tokens packages with lifetime and different levels.
• Relational database: Use of SQLite/Postgres for user data storage and management.
• Vector database: Use of ChromaDB for storing user documents.
• Role system and access control: Ability to control access to various models, including gpt3.5/gpt4.
• Processing of forwarded messages: Understanding of forwarded messages from other users or channels.
• Consideration of user information: The user's information, as well as the current time and date, are factored in when forming responses.
• Automatic model selection: Automatically switches to a model with a 16k context if the context exceeds 4k tokens.
• Feedback collection: With a button in the menu, users can easily contact administrators.
• LRU-based sessions: The bot stores active data (sessions) only for the latest N users.
• Message regeneration: Ability to regenerate messages using different models.
• Integrated power of langchain components: The bot does not rely on agents from there, but uses langchain to work with a large number of document types.
• Docker-native: The project is designed to run in Docker, but it can also be used without it.

Used frameworks: aiogram, langchain, SQLAlchemy, openai, transformers and torch. Full list of dependencies in requirements.txt.

All configs are in resources/.

Running after installing dependencies and setting up is simple: python main.py.

TODO Roadmap

• ✔️ Adaptive truncation of history
• ✔️ Addition of image perception ability (BLIP)
• ✔️ Ability to fully customize all messages
• ✔️ Ability to create a character within a session
• ✔️ Relational user DB (SQLite)
• ✔️ Automatic multi-language interface
• ✔️ System of roles and access to different models (gpt3.5/gpt4)
• ✔️ Understanding forwarded messages from other users or channels
• ✔️ Consideration of user info in the prompt and current time and date
• ✔️ Refactoring to add langchain and new features
• ✔️ Ability to regenerate messages with different models
• ✔️ Revamping of history truncation mechanism (deleting whole messages)
• ✔️ Support for automatic model detection according to context size
• ✔️ User settings in inline keyboard
• ✔️ Full support for Docker, ChromaDB and Postgres
• ✔️ Function-calling support (Working with any number of documents, Website scraping)
• ✔️ Thorough token package, usage limit
• ⏳ Reminder system / user interest tracking
• ⏳ Non-passive conversation (sending trigger messages from time to time)
• ⏳ Ability to understand voice messages
• ⏳ Ability to generate images
• ⏳ Ability to answer with voice messages

Configuration

The main configuration is described in resources/config.json
Characters are configured in resources/personalities.json
All messages can be changed in resources/messages.json

DB Schema