Power Platform Licensing Assistant

A RAG (Retrieval-Augmented Generation) chatbot that answers questions about Microsoft Power Platform licensing using Azure AI Search and Azure OpenAI.

Architecture

User question
     │
     ▼
┌─────────────────────────────────────────────────┐
│  Gradio Web UI  (http://localhost:7860)          │
│  - Model selector (gpt-4o-mini / DeepSeek-V3.2) │
│  - Chat history displayed in browser             │
└───────────────────┬─────────────────────────────┘
                    │
                    ▼
┌─────────────────────────────────────────────────┐
│  Agentic Loop  (bot_respond in rag_chatbot.py)  │
│                                                  │
│  1. Build messages: [system prompt] + history   │
│  2. Call LLM with search tool available          │
│  3. If LLM calls search tool → run search        │
│     → append results → repeat from step 2       │
│  4. When LLM returns final answer → stream it    │
└────────┬────────────────────────┬────────────────┘
         │                        │
         ▼                        ▼
┌─────────────────┐   ┌───────────────────────────┐
│  Azure OpenAI   │   │  Azure AI Search           │
│  (AsyncOpenAI)  │   │  (azure-search-documents)  │
│                 │   │                            │
│  gpt-4o-mini    │   │  Index: Power Platform     │
│  DeepSeek-V3.2  │   │  Licensing Guide chunks    │
└─────────────────┘   └───────────────────────────┘

Logic Flow

1. User submits a question — Gradio appends it to the chat history as {"role": "user", "content": "..."}.

2. Agentic loop starts — The full conversation history plus a system prompt ("You are a Power Platform licensing specialist…") is sent to the LLM with the search_power_platform_licensing tool available.

3. LLM decides to search — The model almost always calls the search tool first. The tool takes a query string, runs a keyword search against the Azure AI Search index, and returns the top-5 matching document chunks concatenated as plain text.

4. Tool result fed back — The search result is appended to the messages as a tool role message. The LLM is called again with the updated context.

5. LLM generates the answer — Once the model has enough retrieved context it produces its final answer, which is streamed token-by-token back to the Gradio UI.

6. Conversation history preserved — The full history list (stored in Gradio's client-side state) is passed on every turn, so the model can reference prior exchanges within the same session.

Key Files

File	Purpose
rag_chatbot.py	Main application — search client, agentic loop, Gradio UI
test_components.py	Smoke tests for LLM and Azure AI Search connectivity
.env	Runtime secrets (not committed) — see .env.example
.env.example	Template showing required environment variables

Environment Variables

Variable	Description
AZURE_OPENAI_BASE_URL	Azure OpenAI endpoint (https://<resource>.openai.azure.com/openai/v1)
AZURE_OPENAI_API_KEY	Azure OpenAI API key
AZURE_SEARCH_ENDPOINT	Azure AI Search service URL
AZURE_SEARCH_API_KEY	Azure AI Search admin/query key
AZURE_SEARCH_INDEX	Name of the search index containing the licensing guide
AZURE_TENANT_ID	Azure tenant ID (used for CLI credential fallback)

Running

# Verify components are working
python test_components.py

# Launch the chatbot
python rag_chatbot.py
# → open http://localhost:7860

Dependencies

• openai — async OpenAI/Azure OpenAI client
• azure-search-documents — Azure AI Search SDK
• gradio — web UI framework
• python-dotenv — loads .env at startup