ElternLeben.de Conversational AI - Hackathon Project

Transforming Digital Family Support in Germany (with Elternleben.de)

This project was developed for the hackathon focused on building a conversational AI system for ElternLeben.de, a non-profit organization dedicated to providing support, guidance, and expert knowledge for parents in Germany. The aim is to help parents create a loving and healthy environment for their children through trusted online counseling, educational content, and personalized programs.

Project Goal

The primary goal of this project was to develop a conversational AI assistant that can effectively support parents by providing information, guidance, and access to ElternLeben.de's services.

Evaluation Criteria

The solution was to be evaluated based on:

1. Conversational Intelligence:
   • Natural dialog flow
   • Context maintenance
   • Appropriate response personalization
   • Accurate content recommendations
2. Service Integration:
   • Seamless transitions from information to services
   • Effective booking/registration flows
   • Proper handling of user information
   • Analytics capabilities

Development Journey & Attempts

This section outlines the iterative approach taken to build the chatbot.

Attempt 1: Simple Echo Bot & Azure Familiarization

• Objective: To understand the fundamentals of the Microsoft Azure Bot Framework and establish a basic deployment pipeline.
• Implementation: A simple echo bot was developed and successfully deployed on Azure.
• Learnings: Gained familiarity with Azure Bot Service and the initial steps for bot deployment.
• Relevant Azure Documentation:
  • Provision and publish a bot
  • Provision an App Service
  • Provision an Azure Bot resource
  • Set up continuous deployment
  • Bot Framework REST API - Authentication

Attempt 2: Building the Knowledge Base - Embeddings for RAG

• Objective: To process and prepare ElternLeben.de's content for a Retrieval Augmented Generation (RAG) system. This involved creating a searchable index of embeddings.
• Implementation:
  • Leveraged code from the Azure Search OpenAI Demo, specifically the prepdocs.py script and its associated filestrategy.py for data processing.
  • Data was split into manageable chunks.
  • Azure OpenAI's text-embedding-ada-002 model was used to compute embeddings for these chunks.
  • The generated embeddings and content chunks were stored in an Azure AI Search index.
• Outcome: Successfully created an embeddings index in Azure AI Search, forming the knowledge base for the RAG solution.

Attempt 3: RAG Solution Integration & Challenges

• Objective: To integrate the RAG pattern with the bot, enabling it to retrieve relevant information from the knowledge base and generate contextual responses using a powerful language model.
• Implementation:
  • The RAG solution was integrated into the bot's backend. Key components involved:
    • app.py and bots.py (main application logic)
    • approaches/ (different RAG strategies and interaction logic)
  • Azure AI Search was used to query the previously created index.
  • Azure AI Services provided access to the gpt-4 model for generating responses based on retrieved context and the text-embedding-ada-002 model for processing user queries if needed.
• Challenge Encountered: Faced OpenAI rate limit issues. An increase in quota is required to fully test and operate the RAG-powered bot at scale. This involves submitting a request to Azure.

Technologies Used

• Cloud Platform: Microsoft Azure
• Bot Framework: Azure Bot Service
• Language Models & Embeddings:
  • Azure OpenAI Service
    • gpt-4 (for generation)
    • text-embedding-ada-002 (for embeddings)
• Search & Data Storage: Azure AI Search (formerly Azure Cognitive Search) for storing and querying embeddings.
• Backend Language: Python

Future Work & Next Steps

• Customize Prompts for ElternLeben.de Data:
  • Modify the prompt files located in app/backend/approaches/prompts/ (e.g., chat_query_rewrite.prompty and chat_answer_question.prompty).
  • Tailor the system messages and few-shot examples within these prompts to specifically address the context of ElternLeben.de's services and user queries, moving away from generic examples (e.g., "Assistant helps the company employees with their healthcare plan questions..." should be changed to something like "Assistant supports parents with questions about child development, parenting challenges, and ElternLeben.de's services...").
• Enhance Service Integration (Booking/Registration):
  • Azure Functions: Develop Azure Functions to wrap calls to mock FastAPI endpoints (for initial development) and later to real booking/registration APIs (e.g., Zoom/SimplyBook). This will centralize error handling, authentication, and logging for these external service interactions.
  • Azure API Management (APIM): Implement APIM to act as a unified gateway for both mock and real service APIs. This will enforce policies like rate limits, CORS, client-credential authentication, and provide usage analytics.
  • Azure Logic Apps: Design visual workflows using Azure Logic Apps to automate post-booking/registration processes. For example, trigger workflows on successful registration to:
    • Send confirmation emails (e.g., via SendGrid).
    • Send SMS reminders (e.g., via Twilio).
    • Push calendar invites (e.g., to Outlook/Google Calendar).
  • Azure Bot Framework Integration: Ensure seamless transitions within the chatbot (managed by Azure Bot Framework) to initiate these booking/registration flows via Azure Functions and APIM.