Lesson · Hinglish · ⏱ 9 min padhne mein

📚 RAG: LLM ko aapke documents se jawaab dene dena

Q: Aapka RAG bot galat jawaab de raha hai. Sahi document maujood hai, lekin kabhi retrieve nahi hota. Pehle kya fix karoge?

Retrieval: chunking, embeddings ya hybrid search. Agar sahi text prompt tak pahunchta hi nahi, to koi LLM sahi jawaab nahi de sakta. Pehle retrieval theek karo.

🏠 Asli zindagi ki misaal

Open-book exam: aap poori textbook ratt nahi lete. Sawaal aane pe sahi pages dhoondhte ho, padhte ho, aur unhi se jawaab likhte ho, chapter ka reference dekar. RAG (Retrieval-Augmented Generation) LLM se aapki company ke documents ke saath open-book exam dilwata hai.

💡 Aasaan shabdon mein

Do phases hain. Indexing: documents ko chunks mein todo (har ek kuch paragraphs ka), har chunk ka embedding banao, aur vector database mein store karo. Answering: user ke sawaal ka embedding banao, sabse relevant chunks retrieve karo, unhe prompt mein daalo ('Sirf in sources se jawaab do…'), aur LLM ko citations ke saath jawaab likhne do. RAG model ko retrain kiye bina jawaab current aur aapke data pe based rakhta hai. RAG ki zyaadatar quality problems retrieval ki hoti hain, jaise galat chunking ya missing documents, isliye retrieval ko final answer se alag naapo.

🎬 Dekho kaise kaam karta hai

Indexing ek baar hoti hai: documents → chunks → embeddings → vector DB. Har sawaal pe sabse acche chunks retrieve karo, prompt mein jodo, aur source ke saath jawaab do.

🪜 Step by step

Documents ikattha karo aur saaf karo: duplicates, purane versions aur bekaar text hatao.
Headings aur paragraphs ke hisaab se chunks banao, har ek kuch sau tokens ka, aur title, section aur date metadata mein rakho.
Chunks embed karke store karo; har sawaal ke liye top matches nikaalo (behtar hai hybrid search plus reranker).
Prompt mein likho: "Answer only from the sources below and cite them. If they don't contain the answer, say so." Yaani sirf sources se jawaab, citation ke saath, aur jawaab na mile to saaf bolo.
Asli sawaalon ka eval set banao; retrieval (sahi chunk mila?) aur answers (sahi aur source ke mutaabik?) alag alag naapo.

🔑 Zaroori shabd

Chunking: Documents ko chhote, search-laayak tukdon mein todna.
Retrieval: Sawaal ke liye sabse relevant chunks dhoondhna.
Grounding / citations: Sirf retrieve kiye sources se jawaab dena aur dikhaana ki kaunse.

👀 Example dekho

python

question = "How many casual leaves do new joiners get?"
chunks = vector_db.search(embed(question), top_k=4)          # retrieve
context = "\n\n".join(f"[{c.source}] {c.text}" for c in chunks)
prompt = f"Answer ONLY from the sources below. Cite them. If not found, say so.\n\n{context}\n\nQ: {question}"
answer = llm(prompt)   # → "New joiners get 8 casual leaves per year [HR-Policy-2026 §4.2]"

Retrieve → prompt mein jodo → generate. Saath mein imaandaar 'nahi mila' rule.

⚠️ Aam galtiyaan jinse bachna hai

Fixed size chunks jo tables aur lists ko beech se kaat dete hain. Document ka structure maano.
Prompt mein 30 chunks thoos dena. Zyaada behtar nahi; kuch relevant chunks aksar bahut saaron se behtar hain.
Permissions ignore karna, jisse users woh documents dekh lete hain jo nahi dekhne chahiye. Retrieval ke time filter karo.

✅ Jaldi se check karo

Aapka RAG bot galat jawaab de raha hai. Sahi document maujood hai, lekin kabhi retrieve nahi hota. Pehle kya fix karoge?

Temperature badhao
Bada LLM le aao
Retrieval: chunking, embeddings ya hybrid search

Jawaab dekho

C. Retrieval: chunking, embeddings ya hybrid search: Agar sahi text prompt tak pahunchta hi nahi, to koi LLM sahi jawaab nahi de sakta. Pehle retrieval theek karo.

🙋 Seekhne wale yeh poochhte hain

RAG ya bada context window?

Bahut bade context windows ke saath bhi RAG aksar sasta, tez aur zyaada accurate hai, kyunki model sirf relevant cheez dekhta hai. Ek chhote document ke liye seedha prompt mein daal do.

Index ko up to date kaise rakhun?

Ek sync job chalao jo badle documents dobara embed kare aur hataye gaye documents delete kare, changes pe ya schedule pe.

📌 Yaad rakho

RAG = apne docs mein search karo, phir unhi se citations ke saath jawaab do.

📖 Is lesson ke shabd (AI dictionary)

Poora course, aapke roadmap ke hisaab se

Progress tracking, quick checks aur step-by-step order ke saath. Yeh lesson in courses ka hissa hai:

Data Engineer (AI focus) · AI Product Engineer · LLM / Agent Engineer · AI Forward Deployed Engineer

Course mein seekho