A full-stack web application leveraging AI to answer new employee questions based on provided company context. Built as a 3-day sprint project.

Live Demo: https://new-employee-assistant.streamlit.app

Onboarding new employees involves answering many common questions about company policies, tools, and culture. This project aims to automate and streamline this process by providing an AI-powered assistant that can answer these questions instantly based on a predefined context document.

This serves as a demonstration of building a full-stack AI application integrating a Large Language Model (LLM), backend API, database, and a simple frontend UI within a tight timeframe.

• Contextual Q&A: Ask questions in natural language about company policies, tech setup, culture, etc.
• AI-Powered Answers: Uses Google Gemini (gemini-1.5-pro) to generate answers based only on the provided internal context document.
• Simple Chat Interface: Clean and easy-to-use UI built with Streamlit.
• Q&A Logging: Questions and their corresponding AI answers are logged to a Supabase (PostgreSQL) database for potential review or future analysis.

• Frontend: Streamlit (UI Framework), Python, Requests (HTTP Client)
• Backend: FastAPI (API Framework), Python, Uvicorn (ASGI Server)
• AI: Google Gemini API (gemini-1.0-pro) via google-generativeai library
• Database: Supabase (Managed PostgreSQL) via supabase-py library
• Deployment:
  • Frontend: Streamlit Cloud
  • Backend: Render
  • Database: Supabase

1. User: Interacts with the Streamlit frontend UI.
2. Frontend (Streamlit Cloud): Captures the user's question and sends an HTTP POST request to the backend API.
3. Backend (Render - FastAPI):
   • Receives the question via the /ask endpoint.
   • Constructs a prompt containing instructions, predefined company context, and the user's question.
   • Sends the prompt to the Google Gemini API (using API key stored securely as an environment variable on Render).
   • Receives the generated answer from Gemini.
   • Logs the question and answer to the Supabase database (using DB URL/Key stored securely as environment variables on Render).
   • Returns the answer as a JSON response to the frontend.
4. Frontend (Streamlit Cloud): Receives the JSON response and displays the AI's answer in the chat interface.

1. Prerequisites:

   • Python 3.10+
   • Git
   • Accounts for: Google AI Studio (for Gemini API Key), Supabase (for DB URL/Key), Render (optional, for deployment reference), Streamlit Cloud (optional, for deployment reference).

2. Clone Repository:

   git clone https://github.com/SebastianMoseres/onboarding-assistant.git
   cd onboarding-assistant

3. Create .env File:

   • In the root onboarding-assistant directory, create a file named .env.
   • Add your secret keys (replace placeholders with your actual keys):

     # .env file
     SUPABASE_URL=your_supabase_project_url_here
     SUPABASE_KEY=your_supabase_service_role_key_here
     GOOGLE_API_KEY=your_google_api_key_here
     

   • Note: This file is listed in .gitignore and should NOT be committed.

4. Setup Backend:

   cd backend
   python -m venv .venv
   source .venv/bin/activate  # Or .\venv\Scripts\activate on Windows
   pip install -r requirements.txt
   # Leave venv active and run the server (in this terminal):
   uvicorn main:app --reload --port 8000

5. Setup Frontend (in a new terminal):

   cd ../frontend # Navigate back to root, then into frontend
   python -m venv .venv
   source .venv/bin/activate  # Or .\venv\Scripts\activate on Windows
   pip install -r requirements.txt
   # Leave venv active and run the app (in this second terminal):
   streamlit run app.py

6. Access the app locally, usually at http://localhost:8501.

• Load Context from Files/DB: Instead of hardcoding COMPANY_CONTEXT, load it from text files, Markdown, or a dedicated database table.
• Vector Database for Context Retrieval: For larger documents, implement a vector database (like Supabase Vector, Pinecone) and retrieval-augmented generation (RAG) to find the most relevant context chunks dynamically before sending to the LLM.
• User Authentication: Add user logins (e.g., using Supabase Auth, Clerk) if different users need different contexts or history.
• More Sophisticated Error Handling: Improve user feedback on errors.
• Streaming Responses: Show the AI answer token-by-token as it's generated for better perceived performance.

Project developed by Sebastian Moseres.