AI, LLM & RAG Exploration Repository

A comprehensive collection of Jupyter Notebooks dedicated to learning, experimenting, and implementing modern AI techniques — from LLM fundamentals to advanced Retrieval-Augmented Generation, Agentic workflows, Fine-Tuning, and LangChain mastery.

📄 The primary dataset used across RAG experiments is the seminal paper "Attention Is All You Need" (NIPS-2017-attention-is-all-you-need-Paper.pdf).

---

📂 Repository Structure

AI/
├── 📓 RAG & Vector Search Notebooks
├── 📓 Agents & Advanced RAG Notebooks
├── 📓 Fine-Tuning & Evaluation Notebooks
├── 📁 LangChain/          ← Full LangChain course (13 notebooks)
├── 📁 Huggingface/        ← HuggingFace NLP Course (6 modules)
├── 📁 Hands-On-Large-Language-Models/  ← LLM internals (5 chapters)
├── 📁 lora_pirate_model/  ← LoRA fine-tune output
├── 📁 pirate_lora_adapter/ ← LoRA adapter weights
└── 📁 archived/           ← Older ML/DL notebooks

---

🧠 RAG & Vector Search

embedding.ipynb

An exploration into text embeddings — how to convert text into high-dimensional vector representations using embedding models. A crucial prerequisite for all Vector RAG pipelines.

RAG.ipynb & RAG_pdf.ipynb — Traditional Vector RAG

Complete pipelines implementing Vector-based RAG over plain text and PDF documents.

• Workflow: Load PDF → Chunk text → Generate embeddings → Store in ChromaDB → Retrieve semantic matches → Generate answers with local LLMs
• Strengths: Broad semantic search and similarity matching

Hybrid_Search_RAG.ipynb — Hybrid Search RAG

Combines traditional keyword-based (BM25) search with semantic vector search to improve retrieval accuracy and reduce failure modes of pure embedding-based retrieval.

Graph_RAG.ipynb — Vector-less Graph RAG

An alternative RAG approach that abandons embeddings in favor of a Knowledge Graph.

• Workflow: LLM extracts (Subject, Relation, Object) triplets → builds a directed graph with networkx → finds exact node matches → retrieves 1-hop subgraph
• Strengths: High precision, 100% explainability, excellent multi-hop reasoning

---

🤖 Agents & Advanced RAG

Function_Calling.ipynb

Foundation for Agentic workflows. Demonstrates how to give an LLM access to external Python functions (tools), enabling it to take actions rather than just generating text.

ReAct_Agent.ipynb

Builds an AI Agent from scratch using the ReAct (Reason + Act) pattern — a while loop lets the LLM think, execute a tool, observe the result, and decide when to stop.

Agentic_RAG_Tutorial.ipynb

Introduces Agentic RAG routing using LangGraph and LlamaIndex Workflows. The LLM acts as a router deciding whether to query a vector database or perform a web search.

Self_RAG.ipynb — Self-Reflective RAG

Advanced pipeline using LangGraph with a built-in "Grader" step that evaluates whether retrieved documents actually answer the question and automatically rewrites the query if retrieval was poor.

---

🏋️ Fine-Tuning & Evaluation

LoRA_PEFT_Tutorial.ipynb

Demonstrates Parameter-Efficient Fine-Tuning (PEFT) using LoRA (Low-Rank Adaptation) and QLoRA for memory-efficient training. Fine-tunes an LLM to adopt a "sarcastic pirate" persona using SFTTrainer.

• Output artifacts: lora_pirate_model/ and pirate_lora_adapter/

RAGAS_Tutorial.ipynb

Evaluates RAG pipelines using the RAGAS framework — programmatically assesses retrieval accuracy, generation quality, and answer relevance using the Gemini API.

---

🦜🔗 LangChain — From Zero to Production (LangChain/)

A complete hands-on course for building LLM-powered applications — 100% free, no OpenAI required. Uses Groq, Google Gemini, Ollama, and HuggingFace.

#	Notebook	Topics
00	00_FREE_Setup_Guide.ipynb	Free providers, API keys, connectivity tests
00	00_LangChain_Overview.ipynb	Architecture, LCEL, core abstractions
01	01_LLMs_and_ChatModels.ipynb	ChatGroq, Gemini, Ollama, streaming, structured output
02	02_Prompt_Templates.ipynb	PromptTemplate, few-shot, CRAFT framework
03	03_Output_Parsers.ipynb	StrOutputParser, JSON, Pydantic, OutputFixingParser
04	04_LCEL_Chains.ipynb	Pipe operator, parallel chains, branching, fallbacks
05	05_Memory_and_Conversations.ipynb	Message history, SQLite persistence, multi-session chatbot
06	06_Document_Loaders.ipynb	PDF, CSV, web, directory loaders; text splitters
07	07_Embeddings_and_VectorStores.ipynb	HuggingFace embeddings, FAISS, Chroma, retrievers
08	08_RAG_Retrieval_Augmented_Generation.ipynb	Full RAG pipeline, citations, conversational RAG
09	09_Tools_and_Agents.ipynb	Custom tools, ReAct agent, tool-calling, agent memory
10	10_LangGraph.ipynb	StateGraph, multi-agent systems, human-in-the-loop
11	11_LangSmith_Observability.ipynb	Tracing, evaluation, custom evaluators, production monitoring
12	12_Production_Best_Practices.ipynb	Retry, caching, async, cost optimization, security

📖 See LangChain/README.md for full module breakdowns and the sequential learning path.

---

🤗 Hugging Face NLP Course (Huggingface/)

Hands-on exploration of the Hugging Face ecosystem, adapted from the official HuggingFace NLP Course.

Module	Content
1_Transformer_Models/	Transformers.ipynb — High-level Transformer architecture overview and pipeline API
2_Using_Transformers/	Behind the pipeline (PyTorch).ipynb, Models (PyTorch).ipynb, Tokenizers.ipynb, attention masking, batching
3_Fine_tuning/	Fine-tune_TrainerAPI.ipynb, full native training loop (section3), learning curve analysis
4_Sharing_Models/	sharing_models.ipynb, using pretrained models from the Hub
5_Dataset_Library/	dataset_lib.ipynb, data_preprocessing.ipynb, dataset_creation.ipynb, Semantic_Search_FAISS.ipynb, bigdata preprocessing
6_Tokenizer_Library/	Byte-Pair_Encoding.ipynb, WordPiece_Tokenization.ipynb, Unigram_tokenization.ipynb, Fast_Tokenizer.ipynb, Fast_Tokenizer_QA.ipynb, training_tokenizer.ipynb

---

📚 Hands-On LLMs (Hands-On-Large-Language-Models/)

Structured walkthrough of core LLM concepts with interactive notebooks:

Chapter	Topic
Chapter 1	Introduction to Language Models
Chapter 2	Tokens and Token Embeddings
Chapter 3	Looking Inside LLMs
Chapter 4	Text Classification
Chapter 5	Text Clustering and Topic Modeling

---

🛠️ Technology Stack

Category	Tools
Local LLMs	Ollama — llama3.2:3b, qwen2.5:7b-instruct
Cloud LLMs (Free)	Groq (llama-3.1-8b-instant, llama-3.3-70b-versatile), Google Gemini Flash
Orchestration	langchain, langgraph, llama-index
Hugging Face	transformers, datasets, evaluate, peft, trl
Vector Database	chromadb, faiss-cpu
Graph Database	networkx (in-memory)
Embeddings	sentence-transformers (all-MiniLM-L6-v2)
Evaluation	ragas

---

🚀 Getting Started

1. Create a virtual environment

python -m venv .venv
# Windows
.venv\Scripts\activate
# macOS/Linux
source .venv/bin/activate

2. Install dependencies

# Core RAG & Agents
pip install langchain langchain-community langchain-ollama langchain-groq langchain-google-genai
pip install chromadb faiss-cpu sentence-transformers pypdf networkx matplotlib jupyter
pip install langgraph llama-index-core llama-index-llms-ollama

# Fine-tuning
pip install transformers datasets evaluate peft trl

# Evaluation
pip install ragas

# LangChain course extras
pip install langchain-huggingface langchain-text-splitters python-dotenv

3. Configure API keys

Create a .env file in the project root (AI/.env):

# Groq — free, fast (recommended)
GROQ_API_KEY=gsk_your_key_here

# Google Gemini — free alternative
GOOGLE_API_KEY=AIza_your_key_here

# LangSmith — optional, for tracing
LANGCHAIN_API_KEY=ls__your_key_here
LANGCHAIN_TRACING_V2=true
LANGCHAIN_PROJECT=ai-exploration

4. Install Ollama (for local notebooks)

# Download from https://ollama.com and then pull models:
ollama run llama3.2:3b
ollama run qwen2.5:7b-instruct

5. Launch Jupyter

jupyter notebook

---

📋 Recommended Learning Path

Embeddings → RAG → Hybrid Search → Graph RAG
                ↓
     Function Calling → ReAct Agent
                ↓
     Agentic RAG → Self RAG
                ↓
     LoRA Fine-Tuning → RAGAS Evaluation
                ↓
     LangChain/ course (00 → 12)
                ↓
     Huggingface/ NLP course (1 → 6)
                ↓
     Hands-On-Large-Language-Models/ (Ch. 1 → 5)