LangChain

LangChain is an open-source framework designed to simplify the development of applications powered by LLMs.

Why Do we need LangChain?

Although LLMs are powerful, they have several limitations:

1. No Long-term Memory: Traditional LLMs process each request independently and they do not remember past conversation
2. No Access to Private Knowledge: LLMs cannot ans the info about company docs etc
3. No Real-Time information
4. Inablity to use External Tools: Tasks like database queries, web searches, and API interactions require additional system, LLMs alone can’t handle it.

Langchain help developers to build applications by combining:

• LLMs
• Prompt Templates
• Memory Systems
• Retrieval Mechanisms
• Vector Databases
• External Tools and APIs
• Agents

Instead of a simple workflow:

User → LLM →Response

Langchain enables:

User →Prompt →LLM →Memory →Tools →Database →Response

Core Components

1. Models: It is the foundation of every LanChain applications. It supports

• OpenAI, Anthropic Claude, Google Gemini, Meta Llama, Open-Source Hugging Face Models
• These models perform tasks such as text generation, summarization, Translation, Question answering etc

Code Example

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(model="gpt-4o-mini",temperature=0.7)

response=llm.invoke("Explain Deep Learning")

print(response.content)

2. Prompt Templates: It allow developers to create reusable and dynamic prompts.

Hardcoded Prompt:

prompt = """

Explain ML

"""

It is not reusable

Prompt Template Example

from langchain.prompts import PromptTemplate

prompt = PromptTemplate(input_variables=["topic"],template="""Explain{topic}""")

prompt.format(topic="NLP")

ChatPromptTemplates: It is used for chat models

Structure:

System Message, Human Message, AI Message

from langchain.prompts import ChatPrompt Template

prompt=ChatPromptTemplate.from_messages(

[( "system","You are an AI Professor"),

("human","Explain {topic}")

])

Types of Prompting:

1. Zero-Shot Prompting: In this, no example has provided in prompt
2. One-Shot Prompting: It has one example
3. Few-Shot Prompting: It has multiple examples
4. Chain of Thought Prompting: It encourages reasoning i.e., think step-by-step.
5. ReAct Prompting: Reason+Action.

3. Chains: It combines multiple operations into a single worlflow. Instead of manually calling each component, LangChain allows developers to connect them into a pipeline.

Get Shraddha Tiwari’s stories in your inbox

Join Medium for free to get updates from this writer.

Remember me for faster sign in

Types of Chain:

1. Simple Chain: Input →Prompt →LLM →Output
2. Sequential Chain: Output of one chain becomes input of another

Input →Summary Generator →Quiz Generator →Output

Example:

chain1 = prompt1|llm

chain2 = prompt2|llm

final_chain = chain1|chain2

3. Parallel Chain: Multiple chains run simultaneously.

Input →(Summary, Sentiment, Keywords) →Combined

It is useful for speed.

4. Output Parsers: LLMs returns text, but Applications often need structured data.

Without Parser:

Name: Kajal

Age: 18

With Parser:

{

“name”:”kajal”,

“age”:18

}

Types

1. StrOutputParser: It returns String

from langchain_core.output_parsers import StrOutputParser

2. JsonOutputParser: Return JSON

from langchain.output_parsers import JsonOutputParser

3. PydanticOutputParser: Returns validated objects

class User(BaseModel):

name:str

age:int

5. Memory: It lets LLMs remember the past conversation memory

Types

1. ConversationBufferMemory: It stores entire chat.
2. ConversationBufferWindowMemory: Stores last N conversations
3. ConversationSummaryMemory: It stores summary, instead of entire chat
4. Vector Memory: It stores conversations as embeddings. Useful for embeddings.

6. Document Loaders: It load external data

PDF: PyPDFLoader, DOCX: Docx2txtLoader, CSV: CSVLoader, Website: WenBaseLoader, Text: TextLoader, JSON: JSONLoader

from langchain_community.document_loaders import PyPDFLoader

loader = PyPDFLoader("notes.pdf")

notes = loader.load()

7. Text Splitters: It break large documents into chunks. LLMs have token limits, So the solution is:

PDF →Chunk1, Chunk2, Chunk3 etc

RecursiveCharacterTextSplitter

splitter = RecursiveCharacterTextSplitter(chunk_size=1000.chunk_overlap=200)

Without chunk_overlap context will be lost.

8. Embeddings: Numerical representation of text, as system can’t understand words

Models:

• OpenAI Embeddings, Sentence Transformers, BGE, E5, Instructor XL