Agentic AI

How to build an Agent

Variables

• Model
  • Foundational Models:
    • OpenAI
    • Google Gemini
    • Anthropic
    • LLAMA
  • Small language models:
    • LLAMA
    • Hugging Face Transformers:
      • microsoft phi
      • Google Gemma
      • …
      • …
  • Fine tuned models
• Framework
  

Workflow steps

• Static Steps: (Router) This is a fixed set of steps and conditional logic around when which node has to be called
  
• Dynamic Steps: (Autonomous Agent): LLM decides steps
• Dynamic steps has very less reliability.
• To exactly solve this problem, Langgraph was introduced.
• Langraph also provides
  • Human in the loop options
  • Persistence
  • Tools
  • LLM
  • State

LangGraph Concepts

• Refer Here for casual reading of graph theory
• Node: This can be any computation
  • interacting with llm
  • calling tools
  • retrieving data from vectorstore
• Edge: Connection between node to node

• State: This is common data that is accessible to all the Nodes when they are executing. Nodes can change the state
  

• Graph is a workflow defined using nodes, edges and state