Project Phases & Development Roadmap

Phase-1: Research Layer

Phase-2: Bedrock models:

someone researched a topic and published a paper, but they dont have funding/money so they cant implement it. here comes some big company from this paper they make a bedrock models.

Phase-3: Providers:

these bedrock model became providers. providers are LLM.

Phase-4:

This is where youn comes in picture

You are going to use these LLM models and create some great AI applications.

Plan of Action:

• 1.LLM foundation
• 2. RAG System
• 3. Agentic AI
• 4. Project and deployment

What is LLM?

LLM= Large Language Model Now break it into 3 words:

1. Large

   • It is trained on a huge amount of text data. (books,websites,wikipedia,articles,code,conversations, etc..).

2. Language

   • It understands and generates human language. Like English,Hindi,Spanish... even programming languages like Python.

3. Model

   • It is a deep learning system (a neural network ) that learns patterns from data.

• When you type : "What is Python?"[converts to number(embeddings)->[0.0,1,1,1.1,1,1,0.2,0.2...]]

The LLM does NOT:

• Search Google
• Think like a Human
• Understand meaning like we do

Instead It:

• Looks at patterns it learned.
• Predicts the next most likely word
• Keeps predicting word by word.

Some Popular LLMs:

• GPT Models
• Gemini Models
• LLaMa
• Claude
• Grok
• Mistral ... etc

How to use these models:

Problem: If every provider has a different SDK, different syntax, different response format... How do we build scalable AI applications?

• LangChain is the solution.

Langchain Components:

1. Models
2. Prompts
3. Chains
4. Memory
5. Indexes
6. Agents

Models:

Models are the core intelligent layer. They generate text,embeddings or other outputs. They connect your application to LLM providers like:

• OpenAI
• Google DeepMind
• Anthropic

In a practical Generative AI, there are mainly 3 important types of models:

1. Chat/ Language Models

• Generate text
• Answer questions
• Write code
• Summarize content

2. Embedding Models

• Convert text into numbers (vectors)
• Used in search & RAG
• Help find similar content

3. MultiModal Models

• Work with more than just text
• Can process images, audio or files

Prompts:

A prompt is the instruction we give to a language model,

It tells the model:

• What to do
• How to respond
• What format to use

Without prompts, the model has no direction.

LLMs are powerful but they are not mind readers. The quality of output depends on:

• How clearly we instruct them.
• Better prompt -> Better output.

Types of Prompts:

1. Simple Prompt: Just a direct instruction. Example: Explain what is Machine Learning. Basic. Straightforward.

2. System + User Prompt Modern chat models use roles:

• System → Sets behavior
• User → Asks question

Example: System: You are a helpful AI teacher. User: Explain LLMs in simple terms. System controls tone and style.

3. Prompt Templates Instead of writing prompts manually every time, we create reusable templates. Example: Explain {topic} in simple terms. Now you can replace {topic} dynamically. This is useful in applications.

4. Structured Prompts Used when we want:

-JSON output -Bullet points -Fixed format

Example: Respond in json format with keys: definition , example This makes output machine readable.

Chains:

• A chain is a sequence of steps connected together. Instead of: One prompt-> One response We create: Step 1->Step 2-> Step 3->final output.

Real AI applications are not one-step. Example: User asks: Summarize this article and translate it into Hindi. What actually happens?

• Summarize text
• Translate summary
• Return result

That's a chain.

Memory:

Memory allows the AI system to remember past interactions. Without memory: Every message is treated independently. With memory: The model can use previous context.

Example: User: My name is Akarsh. Later... User: What is my name? Without memory → Model doesn't know. With memory → It remembers.

Indexes:

Indexes allow us to connect external data to an LLM. LLMs only know what they were trained on. Indexes help them use:

• PDFs
• Documents
• Databases
• Company data
• Websites

LLMs: ❌ Don't know your company data ❌ Don't know private documents ❌ Don't update in real-time So how do we solve this? We connect external knowledge. That system is called Retrieval.

Agents:

Agents are AI systems that can decide what action to take. Unlike normal chains: Chains → Follow fixed steps Agents → Decide the next step dynamically.

Normal LLM: User → Prompt → Response But what if the task requires:

• Searching the web
• Doing calculations
• Calling an API
• Using a database

The LLM alone cannot do that. It needs tools.

virtual env: uv:[https://docs.astral.sh/uv/#__tabbed_1_2] PS D:\Gen-AI> powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex" downloading uv 0.11.8 (x86_64-pc-windows-msvc) installing to C:\Users\Rayied.local\bin
uv.exe
uvx.exe
uvw.exe
everything's installed!

To add C:\Users\Rayied.local\bin to your PATH, either restart your shell or run:

set Path=C:\Users\Rayied\.local\bin;%Path%   (cmd)
$env:Path = "C:\Users\Rayied\.local\bin;$env:Path"   (powershell)

PS D:\Gen-AI> $env:Path = "C:\Users\Rayied.local\bin;$env:Path"

virtual env: uv venv .venv\Scripts\activate

Requirements.txt: langchain langchain-core langchain-community langgraph langchain-openai langchain-google-genai langchain-groq langchain-mistralai python-dotenv faiss-cpu tiktoken fastapi uvicorn requests langchain_huggingface

uv pip install -r requirements.txt