AI Agents aur Tools Guide: LangChain, Memory, AWS Bedrock aur Core Libraries

Introduction

AI ki duniya me ab hum sirf simple prompt dekar jawab lene (chatbots) se aage nikal chuke hain. Aaj ka sabse bada trend hai Autonomous AI Agents — aise smart systems jo user ke goal ko samajh kar khud plan banate hain, jaruri tools (Google Search, SQL DB, APIs) ko chalate hain aur complex tasks ko bina bar-bar human input ke poora karte hain.

Is guide me hum seekhenge:

AI Agents kya hain aur kaise kaam karte hain?
LLM, Tools aur Agents ka aapsi rishta (Relation)
LangChain me Agents aur Tools kaise implement hote hain
Conversation Memory ke types
AWS Bedrock aur Foundation Models
Top AI Softwares & Libraries (LangChain, Hugging Face, Sentence-Transformers, Unstructured, Haystack)

1. AI Agents kya hain? (Agent Fundamentals)

Ek aam LLM (jaise GPT-4) sirf text generate kar sakta hai, wo khud se bahar ki duniya me koi action nahi le sakta. Lekin jab hum LLM ko Tools aur Decision-making Loop de dete hain, tab wo ban jata hai ek AI Agent.

ReAct Framework (Reason + Act)

Agents aam taur par ReAct Pattern par kaam karte hain:

User Goal: "TCS ka stock price check karo aur summary team ko Slack par bhejo"
   │
   ▼
[ Thought ] ──► "Mujhe pehle stock API se live price nikalna chahiye."
   │
   ▼
[ Action  ] ──► Call Tool: get_stock_price("TCS")
   │
   ▼
[ Observe ] ──► Tool Output: "₹3,950 (+1.2%)"
   │
   ▼
[ Thought ] ──► "Ab mere paas data hai, ab mujhe Slack message bhejna hai."
   │
   ▼
[ Action  ] ──► Call Tool: send_slack_message(channel="#finance", msg="...")
   │
   ▼
[ Final Answer ] ──► "TCS ka price check karke Slack par post kar diya gaya hai."

2. Relation: LLM vs Tools vs Agents

In teeno ke beech ka farq samajhne ke liye ye analogy dekhein:

Component	Analogy	Asli Role
LLM	Dimag (Brain)	Language samajhna, plan banana aur text generate karna.
Tools	Hath-Pair (Hands & Legs)	Bahar ki duniya se data lana ya actions execute karna (API, DB, Web search).
Agent	Driver / Coordinator	LLM ki reasoning se decide karna ki kab kaun sa Tool use karna hai.

┌────────────────────────────────────────────────────────┐
│                        AI AGENT                        │
│                                                        │
│   ┌────────────────┐            ┌──────────────────┐   │
│   │   LLM (Brain)  │ ◄────────► │ Memory (History) │   │
│   └───────┬────────┘            └──────────────────┘   │
│           │ Decides tool & params                      │
│           ▼                                            │
│   ┌────────────────────────────────────────────────┐   │
│   │               TOOLS (Shaktiyan)                │   │
│   │  [Google Search]  [SQL DB]  [API Call]  [Code] │   │
│   └────────────────────────────────────────────────┘   │
└────────────────────────────────────────────────────────┘

3. LangChain me Tools aur Agents

LangChain agents aur tools ko aapas me connect karne ke liye ready-made classes provide karta hai.

A. Tools in LangChain

Tools wo functions hote hain jinhe LLM call kar sakta hai.

1. Custom Tool Kaise Banayein (@tool decorator):

from langchain.tools import tool

@tool
def calculate_insurance_premium(age: int, coverage_amount: float) -> float:
    """Calculates yearly insurance premium based on age and coverage."""
    base_rate = 0.02
    if age > 50:
        base_rate = 0.05
    return coverage_amount * base_rate

2. Built-in Tools provided by LangChain:

Search Tools: DuckDuckGoSearchRun, TavilySearch, GoogleSearchAPIWrapper
Database Tools: SQLDatabaseToolkit (Natural language se SQL query run karna)
API & Web Tools: RequestsToolkit (REST APIs call karna), WikipediaQueryRun
Code Execution: PythonREPLTool (Python code likhna aur execute karna)
File System: FileManagementToolkit (Files read/write karna)

4. Conversation Memory (Chat History)

LLMs by default stateless hote hain. LangChain alag-alag memory strategies deta hai:

Memory Strategy	Aasan Matlab	Best Use Case
Buffer Memory	Har message as-it-is save karta hai.	Choti chats ke liye.
Window Memory	Sirf pichle K messages yaad rakhta hai (e.g., last 5).	Token bachane ke liye.
Summary Memory	Purani baaton ki LLM se live summary banwata hai.	Lambi discussions ke liye.
Entity Memory	Specific facts (e.g., User Name = Ahmad, Car = Swift) alag se yaad rakhta hai.	Personalized assistant ke liye.

5. AWS Bedrock Overview

AWS Bedrock Amazon ki ek fully managed service hai jo top AI foundation models ko API ke zariye provide karti hai bina kisi infrastructure ko manage kiye.

AWS Bedrock kyu use karein?

Multiple Foundation Models: Anthropic Claude (Claude 3.5 Sonnet), Meta Llama 3, Amazon Titan, Mistral AI, Cohere sab ek hi jagah.
Enterprise Security: Aapka data AWS VPC ke andar private rehta hai aur public internet par expose nahi hota (HIPAA / GDPR compliant).
Bedrock Agents & Knowledge Bases: Bedrock khud ke agents aur RAG pipelines (Knowledge Bases) support karta hai.

LangChain + AWS Bedrock Example:

from langchain_aws import ChatBedrock

llm = ChatBedrock(
    model_id="anthropic.claude-3-5-sonnet-20240620-v1:0",
    model_kwargs={"temperature": 0.2},
    region_name="us-east-1"
)

response = llm.invoke("Explain AI Agents in one sentence.")
print(response.content)

6. Zaroori AI Softwares & Libraries (Comparison)

AI application develop karte waqt in 5 libraries ka sabse zyada use hota hai:

Software / Library	Kya Hai?	Asli Kaam (Main Use Case)
LangChain	Orchestration Framework	LLMs, Prompts, Tools, aur Chains ko aapas me jodna.
Sentence-Transformers	Python Library (`sbert`)	Local machine par text ko fast aur accurate Embeddings (vectors) me convert karna.
Hugging Face	AI Hub & Ecosystem	Thousands of open-source models (Llama, Mistral), Datasets aur Transformers library ka ghar.
Unstructured	Data Extraction Toolkit	Gande aur complex files (PDFs, Word docs, Scanned images, PPTs) ko saaf text chunks me extract karna.
Haystack (by deepset)	End-to-End RAG Framework	LangChain jaisa alternative framework jo specifically Search, Q&A aur RAG pipelines ke liye optimize hai.

Quick Workflow: Ye sab aapas me kaise milte hain?

[ Raw Complex PDFs ] ──► (1. Unstructured) ──► Clean Text Chunks
                                                    │
                                                    ▼
[ Text Chunks ] ───────► (2. Sentence-Transformers / Hugging Face) ──► Embeddings
                                                                          │
                                                                          ▼
[ Vector DB ] ─────────► (3. LangChain / Haystack Pipeline) ────────► RAG Output
                                                                          │
                                                                          ▼
[ AWS Bedrock / Claude ] ◄── (4. Intelligent Reasoning & Actions) ◄───────┘

7. Quick Revision Summary

Agent = LLM + Tools + Planning Loop (ReAct).
Tools LLM ko calculation, web search aur database access ki shakti dete hain.
Memory chat history maintain karti hai (Buffer, Window, Summary).
AWS Bedrock enterprise-grade cloud service hai jaha Claude aur Llama 3 secure tarike se milte hain.
Unstructured document parse karta hai, Sentence-Transformers embeddings banata hai, aur LangChain/Haystack poore system ko jodte hain.