Agent
An AI Agent is a system that can autonomously break down tasks, make decisions, and execute actions using tools and reasoning.
Agentic Workflow
An agentic workflow is a multi-step autonomous process where an AI system completes tasks without continuous human intervention.
or says: AI auto-complete entire process without human intervention.
e.g.
apply for –> validation –> calculation –> output results
human intervention.
Chunking
Splitting big documents into small pieces so AI can handle them better.
e.g. There are 200 pages in a PDF file, AI cannot read all at once, so splitting file into many small pieces/chunks.
1st piece: 1- 500 words;
2nd piece/chunk: 501 – 1000 words;
3rd piece/chunk: 1001 – 1500 words;
…..
each chunk will become embedding.
It is commonly used in RAG systems to prepare documents for embedding and retrieval.
Cosine Similarity
Cosine similarity measures how similar two vectors are in meaning by comparing their direction in vector space.
or says: A way to measure how similar two pieces of meaning are.
e.g.
apple vs banana : yes, they are very similar.
apple vs car: no, they are not similar at all.
Context Window
Context window is the maximum amount of text an LLM can process at once.
Embedding
Embeddings convert text into numerical vectors that represent meaning.
โappleโ become [0.12, -0.98, 0.33, ……]
โorangeโ become [0.12, -0.98, 0.456, ,,,,,,,] too,
so AI will find
apple = Fruit,
apple != car
or says “simile to a fruit”, and it is not a car. Similar meanings result in closer vector distances, allowing machines to compare semantic similarity instead of exact words.
Fine-tuning
Fine-tuning is the process of further training a pre-trained model on domain-specific data to improve performance in a specialized area.
Hallucination
Hallucination occurs when an LLM generates incorrect or fabricated information while sounding confident.
LangChain
LangChain is a framework for building applications powered by LLMs by connecting models with tools, APIs, and data sources.
in short, chaining interlinkage/link AI , Data, Tools …….
or says “A tool to connect LLMs, data, and tools into applications.”
LangGraph
LangGraph is a framework for building stateful, graph-based AI workflows where agents can loop, branch, and maintain memory across steps.
or says: A workflow system that lets AI follow multi-step flows with loops and decisions.
e.g. SQL Agent,
Write SQL script –> Execute –> Error Alert –> Fix –> Re-try
LLM
Large Language Model. The AI brain that can understand and generate language.
e.g. user asks AI “please write an email”, then output a completed email.
Company uses it to generate Report, analyst Data, auto reply client, …..
MCP
MCP (Model Context Protocol) defines a standardized way for LLMs to interact with external tools, APIs, and data systems.
or says: A standard way for AI to use tools and data systems.
Model Drift
Model drift occurs when a deployed modelโs performance degrades due to changes in real-world data over time.
or says: After the AI was put into use, it started to make mistakes.
why/what’s happened?
maybe, training used old data, now data has changed/updated.
Prompt
A prompt is the instruction given to an LLM.
Well-designed prompts significantly improve the quality and accuracy of model outputs.
e.g.
bad prompt: “write a letter”, — not clearly, what letter you need, thank you letter? complaining letter? ,,,,
good prompt: “Please write a thank you letter to Mary since she gave me a gift.”
Prompt Engineering
Prompt engineering is the practice of designing effective prompts to guide LLM behavior and improve output quality.
or say: Designing better instructions to improve AI responses.
RAG
Retrieval-Augmented Generation. Retrieval-Augmented Generation combines retrieval and generation.
The system first retrieves relevant documents, then uses an LLM to generate an answer based on that information.
e.g. look up HR documents –> pass documents to GPT –> GPT summary then answer question.
Retrieval
Retrieval is the process of searching a knowledge base or vector database to find relevant information before generating an answer.
e.g.
“what is the return policy?” , AI system will look up in “vector DB”, find out “policy document”, pass it to GPT, then answer the question – what is the return policy?
Token
Tokens are the smallest units of text that an LLM processes.
e.g. a sentence like โI love Torontoโ, AI splits โI love Torontoโ into smaller pieces before the model can understand it.
these are tokens,
Token count also determines cost and context limits in LLM systems.
Tool Calling
Tool calling allows LLMs to execute external functions such as APIs, databases, or code to perform real-world actions.
e.g. “AI can “take action”.
>search order
> search database
Vector DB
A database that stores meaning-based vectors for similarity search. It allows AI systems to retrieve semantically relevant documents instead of keyword-based search.
e.g.
there are 10000 file,
HR policy,
IT manual,
Finance report,
……
all of those files become embedding saved in Vector DB. When user asks question, AI will not use “Key-words” to seek, it uses “mean” to match.
Vector Search
Vector search retrieves results based on semantic similarity rather than keyword matching.
or says: Searching by meaning instead of exact words.
