Safety Agent

Converting LangGraph_quickstart into custom safety-agent

Original Workflow Idea

User question
   ↓
Intent detection (is this a statistical question?)
   ↓
If yes → run structured query (e.g., Pandas / SQL / Smartsheet API filter)
If no → normal RAG (semantic retrieval + LLM summarization)
   ↓
Combine the outputs
   ↓
LLM formats the result conversationally

Current implementation

User question ↓ 🧠 Intent detection (statistical vs semantic) ↓ 📊 Statistical Path: 🔍 Semantic Path:

• Azure Blob Storage - Azure AI Search
• Load 12,390 incidents - Vector store query
• LLM generates Python code - Retrieve relevant docs
• Execute Pandas analysis - LLM summarization ↓ ↓ 🤖 LLM formats result conversationally ↓ 🌐 Stream to web interface with chain of thought

Assumptions

• Pipeline exists to extract Smartsheet data
• Smartsheet data loaded into Azure AI search for the RAG vector store and blob storage csv file for Pandas statistical analysiss
• Start developing with a static query, do not develop front end UI
• Output result in the terminal