All About Gen AI Noise

 

AI Developers:

Reference Links:

1. Basic overview of Gen AI, AI agent and Agentic AI:  https://youtube.com/watch?v=O2gerCxEXvc&feature=shared

2. How to select LLm Models to build Gen AI: https://huggingface.co/spaces/lmarena-ai/chatbot-arena-leaderboard

3. When to train code agents with own data. Ideally this is suited for Products based companies who owns huge code base and domain specific codes. Since services companies doesn't own much of code, not suitable for them. Here are the 3 approaches, 

    1. For General Purpose code - use pretrained modesl - GPT-4, CodeLlama

    2. For Doamin specific and consistency with local code and security purpose - Train Model

    3. Hybrid Approach - Use Pretrain model as base.  Fine-tune or embed your codebase into a retrieval-augmented generation (RAG) system. This allows the model to reference your code without full retraining.

When to Use Which?

Scenario	Recommended Approach
You have a massive, stable codebase and budget	Full Training
You want fast, scalable, and cost-effective integration	Hybrid (RAG)
You need offline, private inference	Either, depending on infra
You want to prototype quickly	Hybrid (RAG)

AI Agents with Langraph:

Simple Project: 

https://drive.google.com/file/d/1lO8JIOLekQI3b5-xwHn_x55mwf2meu1_/view?usp=drive_link

Agentic search tool using langraph: https://drive.google.com/file/d/1_mRwJ3aqsH8240NH7H9Qh38N-nN-Ddhk/view?usp=drive_link

Langchain vs Langraph:

When to Use What?

Scenario	Use LangChain	Use LangGraph
Simple chatbot or assistant	✅	❌
Multi-step form or wizard	✅	❌
Dynamic decision trees	❌	✅
Multi-agent collaboration	❌	✅
Stateful, looping workflows	❌	✅
Quick prototyping	✅	❌

Other info:

what is RAG

what is transformer

what is embeddings

what is vectors