Generative AI: Developer's Reference
From LLM fundamentals to production-grade AI systems. Built for interview prep and quick refreshers.
What's Covered
| Topic | Key Concepts |
|---|---|
| LLM Fundamentals | Tokens, Context Window, Temperature, Embeddings, Hallucination |
| Prompt Engineering | Zero-shot, Few-shot, CoT, ReAct, System Prompts, Injection |
| RAG | Indexing, Chunking, Vector DBs, Retrieval, Re-ranking |
| LangChain | LCEL, Prompt Templates, Chains, Memory, Tools, Agents |
| LangGraph | StateGraph, Nodes, Edges, Loops, Supervisor Pattern |
| LangSmith | Tracing, Evaluation, Datasets, Production Monitoring |
The Modern AI Development Stack
┌──────────────────────────────────────────────────────────┐
│ Your Application │
└──────────────────────┬───────────────────────────────────┘
│
┌──────────────────────▼───────────────────────────────────┐
│ Orchestration Layer │
│ LangChain (abstractions) + LangGraph (control flow) │
└──────┬────────────────────────────┬──────────────────────┘
│ │
┌──────▼──────┐ ┌───────▼──────┐
│ LLM APIs │ │ Knowledge │
│ GPT-4 │ │ RAG + Vector│
│ Claude │ │ Databases │
│ Gemini │ └──────────────┘
└─────────────┘
│
┌──────▼──────────────────────────────────────────────────┐
│ Observability: LangSmith │
│ Traces, Evaluations, Datasets, Alerts │
└─────────────────────────────────────────────────────────┘
The Landscape Mindmap
mindmap
root((Generative AI))
Foundation
Large Language Models
Transformers
Tokens
Context Window
Embeddings
Inference Parameters
Temperature
Top-p Top-k
Max tokens
Prompting
Zero-shot
Few-shot
Chain-of-Thought
ReAct
System Prompts
Prompt Injection Attacks
Retrieval Augmented Generation
Document Loading
Chunking Strategies
Embedding Models
Vector Databases
Pinecone Weaviate Chroma
Retrieval Strategies
Re-ranking
Frameworks
LangChain
LCEL Pipe Syntax
Prompt Templates
Chains
Memory
Tools
Agents
LangGraph
StateGraph
Nodes and Edges
Loops and Branching
Multi-agent patterns
LangSmith
Tracing
Evaluation
Datasets
Monitoring
AI Development Shift
Old way: Craft a prompt → call API → done
Modern way: Design systems:
- How does the model access current information? (RAG)
- What tools can it call? (Tool use)
- How does it handle multi-step tasks? (Agents)
- How do you debug when it's wrong? (Observability)
- How do you know it's improving? (Evaluation)
Simple one-shot LLM calls don't need a framework. LangChain, LangGraph, and LangSmith earn their complexity when you're building pipelines, agents, and production systems that need to be debugged, evaluated, and maintained.