AI Agent ek LLM hai jiske paas tools (functions) hote hain. Woh khud decide karta hai ki kaunsa tool chalana hai, result dekhta hai, aur kaam poora hone tak ye loop mein repeat karta hai.
Chatbot = sirf bolta hai. Agent = kaam karke deta hai.
Example: Do dost 📱 Tum poochte ho: "Kal Delhi ja raha hoon, chhata le jaaun?"
Dost 1 (Chatbot): "Pata nahi yaar, October mein aam taur par baarish nahi hoti." (andaaza lagaya)
Dost 2 (Agent): Phone nikaala, weather app khola, dekha "kal baarish ka chance 80%", phir bola: "Haan, chhata le ja!" (khud check karke bataya)
Dost 2 ke paas phone (tool) tha, aur usne khud socha ki weather app kholna hai. Agent bas yahi hai.
Agent ke 4 parts
Part | Kaam |
|---|---|
🧠 Dimaag (LLM) | Sawaal samajhta hai, decide karta hai ki kaunsa tool chalana hai |
🛠️ Haath (Tools) | Normal Python functions jo asli kaam karte hain |
🔁 Loop | Socho → tool chalao → result dekho → kaam poora hone tak repeat |
📝 Memory |
|
Poora flow
Tum: "Kal Jaipur mein garmi kitni hogi?"
↓
LLM: sochta hai → "get_forecast chalao, city=Jaipur, day=tomorrow"
↓
Hamara code: function chalata hai → {"max_c": 34.8, ...}
↓
Result messages list mein judta hai, aur LLM ko wapas jaata hai
↓
LLM: result padhta hai → aur tool chahiye?
├─ Haan → wapas tool call (loop 🔁)
└─ Nahi → final jawab: "Kal Jaipur mein 34.8°C hoga"
↓
Tum: jawab milta hai ✅
🔑 Zaroori baatein
1. LLM khud function nahi chalata.
LLM sirf batata hai ki kaunsa tool, kin parameters ke saath chalana hai. Function hamara code chalata hai.
2. LLM ko tool kya karta hai, ye description se pata chalta hai.
Woh function ka naam, description (docstring) aur parameters padhta hai. Code nahi dekhta. Isliye description saaf likhna zaroori hai.
3. Tool ka output pehle LLM ke paas jaata hai, user ke paas nahi.
Tool kachcha data deta hai, jaise {"max_c": 34.8}. LLM use padh ke achha sa jawab likhta hai.
4. Ek sawaal mein kai tools, ya ek bhi nahi.
"Mumbai aur Shimla mein difference?" → weather tool 2 baar chala, phir calculator.
"Hello" → koi tool nahi chala, LLM ne seedha jawab de diya.
5. Memory: har baar poori history jaati hai.
LLM ko kuch yaad nahi rehta, isliye messages list har baar poori bheji jaati hai. Isi se follow-up sawaal samajh aata hai. Jaise "Kal Shimla?" ke baad "Aur parso?" poocho, to woh samajh jaata hai ki baat Shimla ki ho rahi hai.
6. Loop kab rukta hai?
LLM ne tool nahi maanga aur seedha text diya, to wahi final jawab hai aur loop khatam.
Safety ke liye MAX_STEPS limit hoti hai, jaise 6, taaki loop kabhi atak na jaaye.
7. Error aaye to agent khud sudharta hai.
Tool error de to program crash nahi karta. Error message LLM ko wapas jaata hai, aur woh input badal ke dobara try karta hai.
"Bangalore" → error: "city not found" → LLM ne khud "Bengaluru" se try kiya ✅
8. Control hamare haath mein hai 🔒
Kyunki tool hamara code chalata hai:
Khatarnaak tools (file delete karna, payment) se pehle user se permission le sakte hain.
Tool mein kabhi seedha
eval()mat lagao, warna LLM ka likha koi bhi code tumhare system par chal sakta hai.
🛠️ Tool design ke sabak
Tool simple banao. LLM ko 7 din ki list di, to usne galat din utha liya. Seedha
day="tomorrow"parameter diya, tab sahi jawab aaya.Jo kaam code kar sakta hai, woh LLM se mat karwao. LLM ne likha "using calculator", jabki calculator chalaya hi nahi, aur jod bhi galat kiya. Average aur conversion jaise hisaab tool khud kar ke de.
Prompt badalne se zyada asar tool design badalne ka hota hai.
RAG vs Agent
RAG | Agent | |
|---|---|---|
Kaun decide karta hai | Hamara code (hamesha documents mein search) | LLM khud (kaunsa tool, kab, ya koi nahi) |
Steps | Fix: search → jawab | Jitne zaroori hon (loop) |
Ek line mein: Agent = LLM + Tools + Loop + Memory. LLM sirf decide karta hai, kaam hamara code karta hai.
Chaho to ye version QNA.md ke Q10 mein "Sabse simple samjho" section ki jagah daal doon?