AI aaj sirf ek aisi technology nahi hai jise kuch khaas log hi samajhte hain. Agar aap AI, Machine Learning, Data Science ya AI Engineering mein kaam karna chahte hain, toh kuch basic terms aur concepts ki samajh hona zaroori hai.
Is tutorial mein hum AI se jude 23 important concepts ko step by step samjhenge. Shuruaat AI se karenge aur dheere dheere Neural Networks, LLMs, Transformers, RAG, MCP, AI Agents aur Quantization tak jayenge.
AI, Machine Learning aur Deep Learning
Artificial Intelligence yani AI aisa field hai jisme aise computer systems banaye jaate hain jo un kaamon ko kar saken jinmein aam taur par human intelligence ki zaroorat padti hai. Jaise image recognition, speech translation aur language ko samajhna.
Machine Learning, AI ka ek hissa hai. Ismein machine ko bahut saare data se patterns seekhne diye jaate hain. Machine us data mein maujood patterns ke basis par naye data ke liye prediction kar sakti hai.
Maan lijiye hamare paas lakhon emails hain aur humein pata hai ki kaun si emails spam hain aur kaun si normal hain. Machine Learning algorithm in emails ko dekhkar kuch patterns seekh sakta hai. Baad mein jab koi nayi email aati hai, toh model predict kar sakta hai ki woh spam hai ya nahi.
Deep Learning, Machine Learning ka ek hissa hai jisme Neural Networks ka use kiya jaata hai. Human brain mein neurons ka ek network hota hai. Artificial Neural Networks isi idea se inspired computational networks hain.
Training aur Inference
AI system ko samajhne ke liye do terms bahut zaroori hain, Training aur Inference.
Training ke dauraan model ko bahut saara data diya jaata hai. Model us data mein patterns identify karta hai aur apne parameters ko adjust karta hai. Training ke baad hamare paas ek trained model hota hai.
Iske baad jab model ko naya data diya jaata hai aur woh us data ke basis par prediction karta hai, toh use Inference kehte hain.
Ise ek simple mathematical function ki tarah samajh sakte hain.
Agar input x hai, toh model us input ke basis par output y predict karta hai.
Neural Networks
Neural Network ko samajhne ke liye ek diagram ki kalpana karein jisme bahut saare circles aur unke beech connections hon. Har circle ek neuron ko represent karta hai aur connections ke saath weights jude hote hain.
Ek Neural Network mein aam taur par Input Layer, Hidden Layers aur Output Layer hoti hain.
Input Layer mein data aata hai. Beech ki layers ko Hidden Layers kaha jaata hai aur aakhri layer output deti hai.
Maan lijiye kisi college mein students ke Mid Semester, End Semester aur Class Tests ke marks hain. College ne decide kiya ki Mid Semester ko 40 percent, End Semester ko 40 percent aur Class Test ko 20 percent weight diya jaayega.
Yahan har input ka ek weight hai. Neural Network mein bhi connections ke saath weights hote hain. Ye weights model ko batate hain ki kaunsa input kitna effect daalta hai.
Agar model acche results nahi deta, toh training ke dauraan in weights ko badla jaata hai.
Forward Propagation aur Backpropagation
Jab input Neural Network ki layers se aage ki taraf jaata hai aur aakhir mein prediction nikalti hai, toh ise Forward Propagation kehte hain.
Ab prediction sahi hai ya galat, ye jaanne ke liye Loss calculate kiya jaata hai. Loss batata hai ki model ki prediction actual answer se kitni alag hai.
Iske baad model apne weights ko adjust karta hai. Ye process backward direction mein hoti hai aur ise Backward Propagation ya Backpropagation kehte hain.
Is tarah training mein baar baar Forward Propagation, Loss Calculation aur Backpropagation hota hai. Dheere dheere model apne weights ko better karta jaata hai.
Ek neuron ke andar bhi mathematical calculation hoti hai. Agar teen inputs x1, x2 aur x3 hain aur unke weights w1, w2 aur w3 hain, toh neuron ka basic calculation is tarah likha ja sakta hai:
x1w1 + x2w2 + x3w3 + bias
Iske baad is value par Activation Function apply kiya jaata hai.
Jab Neural Network mein bahut saari layers, neurons aur parameters hote hain, toh use Deep Neural Network kaha jaata hai. Yehi Deep Learning ka base hai.
Large Language Models yani LLMs
Aaj ChatGPT, Gemini aur Claude jaise systems ki wajah se LLM term bahut common ho chuki hai.
LLM ka matlab Large Language Model hai. Ye Deep Learning models hote hain jinhe bahut bade textual datasets par train kiya jaata hai.
Inka kaam human language ko process karna, generate karna, translate karna aur summarize karna ho sakta hai.
LLM ko simple tareeke se samajhne ke liye ek baat yaad rakh sakte hain. LLM text generate karte waqt next token predict karta hai.
Maan lijiye input hai:
Capital of India is
Model prediction kar sakta hai ki iske baad New Delhi se related token aane ki probability zyada hai.
Ye poora process probability aur bahut saari mathematical calculations par based hota hai.
Token aur Tokenization
Computer human language ko bilkul usi tarah directly process nahi karta jaise hum samajhte hain. Isliye text ko chhote units mein divide kiya jaata hai. Inhe Tokens kehte hain.
Token hamesha poora word nahi hota. Ek word kai tokens mein toot sakta hai.
Example ke liye Generative word ko tokenizer alag alag tokens mein divide kar sakta hai. Tokenization ka exact result is baat par depend karta hai ki kaunsa tokenizer use kiya ja raha hai.
Har token ke saath ek Token ID judi hoti hai. Ye ID us particular model ki vocabulary mein us token ko represent karti hai.
Text ko tokens mein badalne ki process ko Tokenization kehte hain.
Context Window
Jab aap kisi LLM ke saath conversation karte hain, toh model ko current conversation ka kuch context diya jaata hai.
Context Window us limit ko represent karti hai ki model ek time mein kitne tokens ke context ko process kar sakta hai.
Is context mein user ke messages, model ke responses, uploaded information aur conversation se judi doosri relevant information shamil ho sakti hai.
Context Window ko working memory ki tarah samajh sakte hain.
Agar conversation bahut lambi ho jaaye, toh context window fill ho sakti hai. Isliye bade tasks ko chhote parts mein divide karna ya purani conversation ko summarize karna useful ho sakta hai.
Badi context window ka matlab hamesha better response nahi hota. Bahut zyada information hone par model ke liye relevant information par focus karna mushkil ho sakta hai.
Vectors aur Embeddings
Token ID se humein token ki identity mil jaati hai, lekin token ka meaning directly samajh mein nahi aata.
Iske liye tokens ko numerical representations mein convert kiya jaata hai jinhe Vectors kaha jaata hai.
Ek vector numbers ki ek series hoti hai. In numbers ke through kisi token ki alag alag properties aur doosre tokens ke saath uske relationships ko represent kiya ja sakta hai.
Maan lijiye hamare paas Cat, Dog aur Lion jaise words hain. Inke vector representations mein related words ek doosre ke kareeb ho sakte hain.
Cat aur Lion ke vectors ek doosre ke kareeb ho sakte hain kyunki dono animals hain. Dog bhi related hai, lekin vector space mein uski position alag ho sakti hai.
LLM ke context mein in vectors ko aksar Embeddings kaha jaata hai.
Text ya tokens ko vector representations mein badalne ki process ko Vectorization kaha jaata hai.
Attention
Attention modern LLMs ka ek bahut important concept hai.
Ek word ka meaning kai baar uske aas paas ke words par depend karta hai.
Example dekhiye:
I code in Python.
A Python bit me.
Pehle sentence mein Python programming language hai. Doosre sentence mein Python snake hai.
Sirf Python token ko dekhkar dono meanings ko alag karna mushkil hai. Iske liye model ko baaki words ka context dekhna padta hai.
Attention mechanism model ko ye samajhne mein help karta hai ki kisi particular token ke liye baaki tokens mein se kaun se tokens zyada relevant hain.
High level par Attention mein Query aur Key jaise vectors ka use kiya jaata hai. Inke beech mathematical operations ke basis par Attention Scores calculate kiye jaate hain.
Agar Python aur code ke beech attention score zyada hai, toh model ko signal milta hai ki yahan Python programming language ke baare mein baat ho rahi hai.
Agar Python aur bit ke beech relation zyada hai, toh model snake wale meaning ki taraf ja sakta hai.
Attention mechanism se juda ek famous research paper 2017 mein publish hua tha, jiska title tha “Attention Is All You Need”.
Transformers
Transformer ek Neural Network architecture hai jo Attention mechanism par based hai.
Aaj ke modern LLMs mein Transformer architecture ka use bahut bade level par kiya jaata hai.
LLM aur Transformer ko ek hi cheez samajhna sahi nahi hai. LLM ek trained language model hai, jabki Transformer woh architecture ho sakta hai jis par model banaya gaya hai.
Transformer ke simplified version ko samjhein toh ismein Attention Block aur Feed Forward Neural Network jaise important parts hote hain.
Input tokens pehle Attention mechanism se guzarte hain. Yahan model context aur tokens ke relationships ko process karta hai. Iske baad information Feed Forward Network se guzarti hai.
Ye process kai layers mein repeat ho sakti hai. Har layer input ki representation ko aur process karti hai.
Jaise jaise layers badhti hain, model input ke zyada complex relationships ko represent kar sakta hai.
Transformer mein attention calculation ki wajah se token count badhne par computation bhi kaafi badh sakti hai. Isi wajah se doosre architecture approaches jaise State Space Models par bhi research aur development hua hai.
Fine Tuning
Har general purpose model kisi particular industry ya task ke liye perfect nahi hota.
Maan lijiye humein medical field ke liye specialized model chahiye. Iske liye ek existing pretrained model liya ja sakta hai aur use medical data par further train kiya ja sakta hai.
Is process ko Fine Tuning kehte hain.
Fine Tuning mein model ko scratch se train karne ki zaroorat nahi hoti. Ek pretrained base model liya jaata hai aur specialized dataset ke basis par use further train kiya jaata hai.
Isse industry specific, task specific ya particular response style wale models banaye ja sakte hain.
Self Supervised Learning
Machine Learning mein data ko broadly labeled aur unlabeled data mein dekha ja sakta hai.
Labeled data mein input ke saath target output bhi diya hota hai. Jaise kisi email ke saath label ho ki woh Spam hai ya Not Spam.
Lekin LLMs ko bahut bade amounts of text par train kiya jaata hai. Itne bade data ko manually label karna practical nahi hai.
Yahin Self Supervised Learning ka idea kaam aata hai.
Model apne existing data se hi training targets bana sakta hai.
Maan lijiye sentence hai:
Don’t judge a book by its …
Model ko kuch part hide karke aage aane wale token ko predict karne ke liye kaha ja sakta hai.
Agar model sahi token predict karta hai, toh prediction achhi hai. Agar galat karta hai, toh loss ke basis par model apne parameters ko adjust karta hai.
Is tarah unlabeled data se training possible hoti hai.
Reinforcement Learning aur RLHF
Reinforcement Learning mein model ko actions lene aur unke results se seekhne ka tareeka diya jaata hai.
Ise ek simple example se samajhiye. Agar koi system chess khel raha hai aur aisa action leta hai jisse use fayda hota hai, toh use positive reward mil sakta hai. Agar action kharab hai, toh penalty mil sakti hai.
Model ka objective reward ko maximize karna hota hai.
Reinforcement Learning ka use games, robotics aur AI systems mein kiya ja sakta hai.
Reinforcement Learning from Human Feedback yani RLHF mein human feedback bhi training loop ka part ban sakta hai.
Example ke liye model do alag responses generate karta hai aur human batata hai ki kaunsa response zyada useful hai. Is feedback ke basis par model ko positive ya negative signal diya ja sakta hai.
RAG yani Retrieval Augmented Generation
RAG ka poora naam Retrieval Augmented Generation hai.
Normal LLM apne training data ke basis par response generate karta hai. Lekin kisi company ke internal documents ya private information ke baare mein model ko information nahi ho sakti.
Maan lijiye kisi company ke paas apne products ke internal documents hain. Company chahti hai ki customer support AI un documents ke basis par answer de.
Iske liye RAG pipeline banayi ja sakti hai.
Is setup mein user query aati hai. System relevant documents ya document chunks ko retrieve karta hai. Phir retrieved information ko LLM ke context ke saath bheja jaata hai aur LLM us information ke basis par response generate karta hai.
Ismein LLM generation kar raha hai aur relevant information retrieval system se aa rahi hai. Isi wajah se ise Retrieval Augmented Generation kaha jaata hai.
RAG mein PDFs, documents, images aur doosre types ka data source ho sakta hai.
Vector Database
RAG pipelines mein Vector Database ka kaafi use hota hai.
Normal keyword search mein exact words match karna zaroori ho sakta hai. Lekin Vector Database semantic search kar sakta hai, yani meaning ke basis par relevant information search kar sakta hai.
Maan lijiye user search karta hai:
Jogging shoes
Database mein koi product “Lightweight Footwear” ke naam se maujood hai. Usmein jogging ya shoes words shayad na hon, lekin uska meaning query se related ho sakta hai.
Vector search aise relationship ko pakad sakta hai.
RAG pipeline mein documents ko pehle chhote chunks mein divide kiya jaata hai. Phir in chunks ke embeddings banaye jaate hain. Original chunks aur unke vector representations ko Vector Database mein store kiya jaata hai.
Jab user query aati hai, query ka bhi vector representation banaya ja sakta hai. Phir database similar vectors search karke relevant chunks return karta hai.
Iske baad ye chunks LLM ko diye jaate hain.
Kuch popular vector database technologies ke examples Pinecone aur Chroma hain.
Few Shot Prompting
Prompting mein hum LLM ko sirf ek question nahi dete. Zaroorat padne par hum examples bhi de sakte hain.
Agar hum chahte hain ki LLM kisi particular style mein response de, toh prompt ke andar kuch example inputs aur desired outputs diye ja sakte hain.
Ise Few Shot Prompting kehte hain.
Agar koi example nahi diya jaata, toh use Zero Shot Prompting kaha ja sakta hai.
Agar bahut saare examples diye jaate hain, toh use Many Shot Prompting kaha ja sakta hai.
Few Shot Prompting RAG systems aur doosre AI applications mein bhi use kiya ja sakta hai.
MCP yani Model Context Protocol
MCP ka poora naam Model Context Protocol hai.
LLM ko kai baar external tools aur external resources se information chahiye hoti hai. Example ke liye kisi GitHub repository se data lena ya customer support platform se ticket information lena.
MCP aise systems ke beech communication ke liye standardized protocol provide karta hai.
High level par MCP setup mein MCP Client aur MCP Server jaise components hote hain.
Maan lijiye LLM ko GitHub repository se information chahiye. MCP Client relevant MCP Server se connect kar sakta hai aur wahan se required information lekar LLM ko de sakta hai.
Isi tarah customer support system ke liye MCP Server kisi support platform se relevant ticket information la sakta hai.
Iska basic purpose LLM ko external tools aur resources ke saath structured tareeke se connect karne mein help karna hai.
Context Engineering
Jab hum LLM ko sirf user query nahi dete, balki uske saath extra relevant information bhi dete hain, toh is poore process ko Context Engineering ke perspective se dekha ja sakta hai.
RAG mein retrieved documents context ka part ban sakte hain.
Few Shot Prompting mein diye gaye examples context ka part ho sakte hain.
MCP se milne wali external information bhi context ka part ho sakti hai.
User preferences, conversation history aur specific instructions bhi context mein shamil kiye ja sakte hain.
Lambi conversation mein context ko manage karne ke liye purane messages ka summary bhi banaya ja sakta hai. Isse poori purani conversation ko har baar context mein bhejne ki zaroorat kam ho sakti hai.
Multimodal Models
Shuruaati language models mainly text ke saath kaam karte the. Lekin modern AI models kai alag forms of data ke saath kaam kar sakte hain.
Aise models ko Multimodal Models kaha jaata hai.
Ye text ko process kar sakte hain. Kuch models images ko samajh ya generate kar sakte hain. Kuch audio aur video ke saath bhi kaam kar sakte hain.
Isliye modality ka matlab us form se hai jisme data maujood hai.
Agar task sirf text based hai, toh text focused model kaafi ho sakta hai. Lekin agar task mein images, audio ya video bhi shamil hain, toh multimodal capabilities useful ho sakti hain.
AI Agents
Ek normal LLM aam taur par query lekar response generate karta hai.
Lekin kai tasks mein humein sirf answer nahi chahiye. Humein koi action bhi karwana hota hai.
Example ke liye GitHub par code push karna, commit create karna, ticket update karna ya kisi external service se information lena.
Yahin AI Agents ka concept aata hai.
AI Agent ko ek aise system ki tarah samajh sakte hain jise koi goal ya task diya jaata hai aur jiske paas kuch tools aur resources ka access hota hai.
Agent us goal ko chhote tasks mein divide kar sakta hai aur available tools ka use karke task execute kar sakta hai.
Agent ke andar LLM reasoning aur planning ka kaam kar sakta hai. Uske aas paas tools, memory, external services aur doosre components ka poora system ban sakta hai.
Example ke liye ek AI Engineering Agent GitHub par code changes kar sakta hai, Jira tickets ke saath kaam kar sakta hai aur Slack par team members ke saath communicate kar sakta hai.
Chain of Thought aur Reasoning Models
Kuch LLMs directly final answer dete hain. Kuch systems kisi complex problem ko kai reasoning steps mein process karte hain.
Kisi problem ko chhote reasoning steps mein break karke answer tak pahunchne ke approach ko Chain of Thought kaha jaata hai.
Aisa reasoning approach complex problems ko solve karne mein help kar sakta hai.
Reasoning Models aise models hote hain jinhe complex problems par step by step reasoning aur planning ke liye banaya jaata hai.
Reasoning ke doosre approaches bhi ho sakte hain, jaise Tree of Thought aur Graph of Thought.
Tree of Thought mein model alag alag possible branches explore kar sakta hai. Agar kisi branch par useful result nahi milta, toh doosre path ko explore kiya ja sakta hai.
Ek model ek saath kai capabilities bhi rakh sakta hai. Example ke liye koi model LLM hone ke saath reasoning model aur multimodal model bhi ho sakta hai.
Small Language Models yani SLMs
Large Language Models ke comparison mein Small Language Models mein parameters ki sankhya kam hoti hai.
Kam parameters ki wajah se inka computation aur inference cost kam ho sakta hai.
Small Language Models ko particular use cases ke liye specialized banaya ja sakta hai.
Example ke liye kisi hospital ke customer support system ke liye ek specialized model banaya ja sakta hai. Kisi legal company ki internal policies ke liye bhi specialized small model use kiya ja sakta hai.
Large models ko general purpose models ki tarah aur small models ko specific use cases ke liye specialized models ki tarah samajh sakte hain.
Distillation aur Quantization
Small Language Model banane ka ek tareeka Knowledge Distillation hai.
Ismein ek Large Language Model ko Teacher Model aur chhote model ko Student Model ki tarah use kiya jaata hai.
Ek hi input dono models ko diya ja sakta hai. Teacher model output generate karta hai. Student model ko teacher ke output se seekhne ke liye train kiya jaata hai.
Is process mein student model teacher model ke behavior aur outputs ko approximate karne ki koshish karta hai.
Is tarah bade model ki capabilities ka kuch part chhote model mein transfer karne ki koshish ki jaati hai.
Ab Quantization ko samajhte hain.
Modern Neural Networks mein bahut saare parameters hote hain aur har parameter numerical value ke form mein store hota hai.
In values ko alag alag numerical precision mein store kiya ja sakta hai, jaise 32 bit, 16 bit, 8 bit ya 4 bit.
Jitni zyada bits use hongi, utni zyada memory ki zaroorat ho sakti hai.
Quantization mein model weights ko lower precision format mein represent kiya jaata hai. Isse memory requirement aur inference cost kam ho sakti hai.
Example ke liye 32 bit weights ko lower precision mein convert karne se model ka memory footprint kaafi kam ho sakta hai.
Quantization ke do common approaches hain, Post Training Quantization aur Quantization Aware Training.
Post Training Quantization mein model ko pehle normal tareeke se train kiya jaata hai aur training ke baad weights ko quantize kiya jaata hai.
Quantization Aware Training mein training ke dauraan hi low precision behavior ko dhyan mein rakha jaata hai. Isse training process zyada complex ho sakta hai, lekin final model ki accuracy ko better tareeke se preserve karne mein help mil sakti hai.
In 23 concepts ko ek saath dekhein toh AI ki kai basic ideas ek doosre se judi hui hain.
AI ek broad field hai. Machine Learning AI ka ek hissa hai aur Deep Learning Machine Learning ka hissa hai.
Neural Networks Deep Learning ka base hain. LLMs language related tasks ke liye bade Deep Learning models hain. Transformers modern LLMs mein use hone wali ek important architecture hain aur Attention Transformer architecture ka ek bada part hai.
Tokens text ko process karne ki basic units hain. Embeddings tokens ko numerical representations dete hain. Context Window model ko current task ki relevant information available karati hai.
RAG external documents se relevant information retrieve karke LLM generation ko support karta hai. Vector Database semantic search mein help karta hai. Few Shot Prompting examples ke through model ko desired response ka pattern dikhata hai.
MCP LLMs ko external tools aur resources se connect karne ka structured tareeka deta hai. Context Engineering in alag alag sources se milne wali relevant information ko manage karne par focus karta hai.
AI Agents LLMs ko tools aur actions ke saath connect karke aise systems bana sakte hain jo sirf answer dene ke bajay tasks execute kar saken.
Reasoning Models complex problems par reasoning aur planning kar sakte hain. Small Language Models kam resources mein specialized tasks ke liye use kiye ja sakte hain. Distillation bade model se chhote model ko train karne ka tareeka hai aur Quantization model ko lower precision mein represent karke memory aur computation requirements ko kam karne ka tareeka hai.
In concepts ko samajhne ke baad AI se judi bahut si discussions ko samajhna aasaan ho jaata hai. Har concept ka mathematical implementation abhi se yaad hona zaroori nahi hai. Pehle ye samajhna zaroori hai ki kaunsa concept kis problem ko solve karta hai aur doosre concepts ke saath uska connection kya hai.
FAQs
Everything you need to know about PromptDon
AI yani Artificial Intelligence aisi technology hai jisme computer systems aise tasks perform kar sakte hain jinmein human intelligence ki zaroorat padti hai, jaise language samajhna, image recognition aur translation.
AI ek broad field hai, jabki Machine Learning AI ka ek part hai. Machine Learning mein system data se patterns seekhta hai aur un patterns ke basis par naye data ke liye predictions karta hai.
Deep Learning, Machine Learning ka ek part hai jisme Neural Networks ka use kiya jaata hai. Deep Neural Networks mein multiple layers ho sakti hain jo data ko process karti hain.
LLM ka matlab Large Language Model hai. Ye Deep Learning models hote hain jinhe bahut bade text datasets par train kiya jaata hai. Ye text ko samajhne, generate karne, translate karne aur summarize karne jaise tasks kar sakte hain.
Token text ka ek chhota unit hota hai. Ek word kabhi ek token ho sakta hai aur kabhi multiple tokens mein divide ho sakta hai. Text ko tokens mein convert karne ki process ko Tokenization kehte hain.
RAG ka matlab Retrieval Augmented Generation hai. Ismein system relevant external information retrieve karta hai aur us information ko LLM ke context mein dekar response generate karwata hai.
Vector Database aise numerical vector representations ko store aur search karne mein help karta hai. Iska use semantic search aur RAG systems mein kiya ja sakta hai.
Few Shot Prompting mein LLM ko prompt ke andar kuch examples diye jaate hain, taaki model desired response ka pattern samajh sake.
Reasoning Model aise AI models hote hain jo complex problems ko multiple reasoning steps ke through process karne ke liye design kiye jaate hain.
Small Language Model yani SLM ek comparatively chhota language model hota hai jisme parameters kam hote hain. Is wajah se kuch use cases mein ise kam computation aur resources ke saath run kiya ja sakta hai.
Knowledge Distillation mein ek bade Teacher Model ke outputs ya behavior se ek chhote Student Model ko train kiya jaata hai. Iska purpose bade model ki capabilities ka kuch part chhote model mein transfer karna hota hai.
Quantization mein model ke weights ko lower numerical precision mein represent kiya jaata hai, jaise 32 bit se 8 bit ya 4 bit. Isse model ki memory requirement aur inference cost kam ho sakti hai.
