Skip to content

Latest commit

 

History

History
249 lines (154 loc) · 9.72 KB

README.md

File metadata and controls

249 lines (154 loc) · 9.72 KB
Raw

Wendy | Mental Health AI Telegram Chatbot

Header

🤝 Participants

  • Vladislav Artiukhov
  • German Bagdasaryan

🎉 Awards

🥇 The solution has won the "AI-Powered Mental Wellness Support Chatbot" hackathon!

The project succeeded in securing 4/9 nominations:

💻 Code Quality
🔄 Continuous Engagement and Support
🔒 Security and Privacy Assurance
🏗️ Architecture Scalability

We got published on the hackathon's winners page!

🌟 Idea

Wendy is an AI Telegram chatbot that has a personality of a helpful and loving pal designed to help her peers with mental health issues.

The prompts are designed to make the AI behave like a real person, creating high engagement for continued chatting with Wendy (implemented).

Besides the conversation with the AI, the application allows users fill out a personal questionnaire that enriches the knowledge of the AI about her partner (implemented).

Users can set up custom notifications, enabling the AI to initiate conversations based on recent changes in users' lives reflected in previous messages (proposal). This brings AI a proactive attitude in conversations.

In case of emotional deprivation, AI may suggest the contacts of local mental health support affiliations based on user's country of residence. Besides, users may access these contacts in their accounts (implemented via Telegram webapps) (proposed).

▶️ Demos

  1. Presentation Defense: link.
  2. Demo for Hackathon: link.
  3. Extended Demo (a real prolonged conversation): link.
  4. Live stream with Nominations & winner selection: link.

Start of the conversation:

Cold conversation start feature is turned off, as it requires a filled out questionnaire and a server hosted on a publicly available server (as far as we know, it is a limitation of the Telegram webapps):

Talking about Harry Potter:

The prompt contained information that Wendy loves reading fantasy books, such as "Harry Potter" and "Lord of the Rings".

Talking about intimate relationship:

To our surprise, the LLM created and developed the narrative that was not inserted in the prompt: Wendy previously had intimate relationship with a guy named Alex. They broke up because of having too little time to spend together.

📝 Project Description

🚀 Hackathon

The project was implemented during the 3-day hackathon "AI-Powered Mental Wellness Support Chatbot" (27.06 — 30.06, 2024) held by Hackathon Raptors.

The main hackathon objectives were to create empathetic AI-user interactions and sustained user engagement, positioning the bot as a bridge to public mental health affiliations.

⚙️ Technical features

  • Cold conversation start: post questionnaire completion, AI always initiates the first chat message; leveraging the Few Shot Prompting technique, we create sample conversation between Wendy and the user (based on the questionnaire). This sample is then used for the AI's second prompt to initiate the conversation.
  • Message summaries: crafting summaries of chat messages with AI and using them in chat history to preserve more details about past conversations1 (implemented).
  • AES encryption: encrypt messages and their summaries before storing them in the database (implemented).
  • Telegram webapp integration: features, other that chatting, are delivered in a stand-alone web app that is attached to the Wendy Telegram bot as a Telegram webapp (implemented).

1 Chat history consists of a concatenation of summarized and normal chat messages (the exact history structure is defined by a selected history crafting strategy).

🧩 Functional features

  • Engaging conversation with AI: utilize GPT-4o and research-based prompting techniques1 to make the AI as human-like as possible.
  • Periodic and occasional notifications: user may select time periods when they would love to hear from Wendy and share their thoughts; besides, Wendy may write occasional messages such as happy birthday, New Year, Christmas or users' country local events greetings (proposed).
  • Personalized Questionnaires: users provide additional information about them to have better more personal dialogs (implemented).
  • Mood classification: Periodically classify user mood based on chat messages via AI to determine when to suggest contacts for mental health organizations.
  • Contacts of local mental health affiliations: contacts are defined based on user's country of residence and are always present in the Telegram webapp or suggested by Wendy (proposed).
  • Mental Health Surveys: users may find standardized self-surveys, such as Beck and Children's Depression Inventories, delivered as interactive forms within the Telegram webapp (proposed).

1 We utilized Few-Short Prompting2 and Emotional Prompting3 techniques that are empirically shown to yield better LLM responses.

2 “Language Models are Few-Shot Learners”, Brown et al, 2020.

3 “Large Language Models Understand and Can be Enhanced by Emotional Stimuli”, Li et al., 2023.

🛠️ Technical details

Stack

TechnologyRationale
TypeScriptTypes + rapid development + npm
Telegram Bot APIPopular message + advanced APIs
OpenAI APIRobust API + advanced LLMs
PostgreSQLPopular relational database
PrismaBest ORM for TypeScript
DockerDeployment + containerization
VitestRapid unit testing

Advanced History Crafting

advanced-history-crafting
The history is a sliding window over the sequence of messages. We greedily try to collect N messages splitting them in halves, where messages from the 1st half (older messages) come as summarized messages, and from the 2nd half as normal ones. It allows to incorporate more information about the past conversations. Messages Summaries are created by an LLM.

An example of a charted history to the LLM:

### Recent updates:
Summarized version of the message:
[Sent by]: user
[Time]: 12 Oct, 2024, 21:00:00
[Summary]: "Tell me about yourself"

Summarized version of the message:
[Sent by]: assistant
[Time]: 12 Oct, 2024, 21:00:30
[Summary]: "I’m Wendy, I love fantasy."


### Last conversation messages:
[Sent by]: user
[Time]: 12 Oct, 2024, 21:00:55
[Message]: "please do not write so much text, it is annoying.."

[Sent by]: assistant
[Time]: 12 Oct, 2024, 21:01:12
[Message]: "I'm sorry about that, Vlad. I'll keep it short."


### Last message:
Hi!

Prompt templates

Prompts are implemented in a way to support dynamic content insertion, allowing use of pre-defined variables. See PromptTemplate class for details.

Questionnaire

Personal questionnaires are implemented as a web form within the Telegram webapp:

questionnaire

Once a user sends the form, we store the data in the database for further use in the system prompt for the LLM.

👨‍💻 Development

Dependencies

  • Docker & docker-compose
  • nodejs & npm

To install the project dependencies run:

make install
# runs: npm install

Environment Variables

  1. Create an .env file in the root of the project and insert the following variables:

    # Database:
# used by `docker/docker-compose.yaml`
POSTGRES_USER=user 
POSTGRES_PASSWORD=pass
POSTGRES_DB=db_name
POSTGRES_PORT=5432
# used by Prisma
DATABASE_URL="postgresql://user:pass@localhost:5432/db_name?schema=public"

# Telegram bot:
BOT_TOKEN=1234567:AAAAAA-BBBBBBBBBBB-CCCCCCCCCCCC

# OpenAI:
OPENAI_API_KEY=sdf_fds-fsdasdfasdfasdfsnfsdfsdfadf

# Encryption (optional: if absent no encryption applied):
ENCRYPTION_KEY=my-encryption-key

  2. To boot up the postgres database in a docker run:

    make compose
# runs: docker-compose --env-file ./.env -f docker/docker-compose.yaml up -d

Server (Telegram Bot)

To start the server run:

make run
# runs: npx ts-node ./src/index.ts