Below are the courses related to building agents along with my notes.
AI Agents in LangGraph, by Langchain, Tavily, DeepLearning.AI - [short course]
- Build an Agent from scrach
- Based on: A simple Python implementation of the ReAct pattern for LLMs, by Simon Willison - blog post
- ReAct Pattern: Stands for “reasoning plus acting.”
- Initial Thought: LLM begins by thinking about what action to take
- Action Execution: The action is executed within an environment.
- Observation and Feedback: Following the action, an observation is returned from the environment.
- Repetition and Refinement: The LLM processes the observation, rethinks, and decides on the next action.
- Continuation: This cycle continues until the LLM decides no further actions are necessary.
- LangGraph components
- Other frameworks:
- Prompt templates for ReAct source
- LangGraph have predifined tools, such as:
from langchain_community.tools.tavily_search import TavilySearchResults
- LangGraph components:
- Nodes: Agents or Functions
- Edges: Connect Nodes
- conditional edges: decision
- Entry point: starting node
- End node (ex:
from langgraph.graph import END)
AgentState: Annotated list of messege that we’ll add to over time
- Agentic Search Tools
- Tavily
- Example of How: break query, to sub queries, Retrieve from different souce, score and filtering, then return top k
- example:
from tavily import TavilyClient
- Persistance and streaming
- Human in the loop