technonotes-hacker

Posted on Nov 4

Promotion - Framework - LangChain

#ai #opensource #llm #programming

Intro

Open-Source Framework
Its used for building the applications powered by LLM.
In simple words --> Langchain helps to connect an LLM with prompts, data , Tools , memory , APIs and other components to build useful AI Application.
LLM can answer your question but you can't build a APLLICATION.

Without LangChain

User --> Prompt --> LLM --> Answer

With Langchain

Why we need Langchain ?

It helps to build real world applications around LLMs by connecting with Prompts, documents , database , API's , Tools , Memory and Other Components.
See LLM is also powerful, but its mainly a TEXT GENERATION Engine.
But for PRODUCTION AI APPLICATION you need entire workflow around it.
Example:

Langchain Architecture

Its a framework and supports many things.
Its like a SET OF COMPONENTS that sits between user/Application & LLM , connecting the model to data, tools and workflow.

Langchain Components

It provides different building blocks with that you can combine to create LLM applications like:

Chatbots
RAG Systems
AI Agents
AI Data Analysts

6 Areas that Langchain is designed to HELP.

Model I/O

Main area of Langchain.
It handles the interaction between the application and the LLM or chat model.
Model I/O = Input --> Model --> Output

Main parts of Model I/O:

1. Prompt Template
2. Models
3. Output Parser

Prompt --> Instruction to the LLM model so that it can perform the task for us.
Instead of repeating the prompt ( hardcoding prompts ) , Langchain allows us to create a reusable templates.
If you want to ask 100 different topics like:

1. Explain Python
2. Explain SQL
3. Explain Deep Learning
........................
........................

You can create a reusable temple which will allow to explain in simple terms.

Explain { Topic } in simple terms.

Models

Model is responsible for generating the response.
Langchain will work with different types of models.
Its supports Multi Model Systems.

StrOutputParser --> It will help to put in Plain TEXT.

Temperature --> Output generation depends on this parameter. In other words , how creative you are going to get the response.

Data Connections

Connecting the LLM Application to the external or private data source so the LLM can retrieve and use information that is not available in its training DATA.
Like Confidential Data , Policy , etc
Eg., A company has 500 PDF documents and we want to build a chatbot that answers questions from the PDFs.

Document Loader is supported by LangChain.
Eg., PDF , CSV , File , Txt , Word Docs , Website , Database , Notion , Google Drive , APIs.
Directly அந்த documentல இருந்து எதுமே எடுக்க முடியாது, அதனால its divided into Text Splitters.
Text Splitters --> Chunks or Tokens , Large documents are divided into smaller chunks.

Now the models can't understand the text data , so EMBEDDINGs coming into picture.
Embeddings --> Converting text into numerical vectors using Semantic or Similarities.
Once the numerical Vectors are created then it needs to be stored.
Vector Store or Vectore Database --> The embeddings are stored in a vector Store or Vectore Database. ( Eg., FAISS , Chroma , Pinecone , Weaviate , Milvus , Qdrant )
Retreiver --> Searches the vector store for information relevant to the user's questions.
RAG --> All these components together are commonly called as RAG - Retrieval Augmented Generation.

Chains in LangChain

Sequence of steps , where the output of 1 step becomes the input of next step.
Chain = Connecting Multiple Operations together to complete a TASK.

Why use CHAINS ?

Reuse workflows
Modularize AI applications
Connect Multiple components Together
Build complex applications like Chat-bots, RAG systems and Agents.

Basic LLM Chain

Takes the user input , puts that into a prompt , sends the prompt to LLM such as Gemini and returns the response.

There are 3 components:

Prompt
LLM
Output Parser

Who is responsible to combine all 3 ? Chain.

Temperature --> Controls the randomness or creativity of the LLM output.

Sequential Chain in LangChain

Executing multiple steps one after the another.
Where the output of one step is passed to the next step.

Connect both the chain , which is called Runnable Pass through in LangChain.
LCEL --> LangChain Expression Language. ( Its a syntax and framework in LangChain used to combine different components into chains using " | ". )

Parallel Chain

Runs multiple independent chains at the same time using the same input.

All 3 chains can work independently.
Example: If we type --> Machine Learning , we want Gemini to generate , Definition , Advantages & Applications.

LangChain Memory

Allows application to remember information from previous interactions and use that information in future conversations.
Without Memory:

User : My name is Leo
AI : Nice to meet you Leo

User : What is my name ?
AI : I don't know

With Memory:

User : My name is Leo
AI : Nice to meet you Leo

User : What is my name ?
AI : Your name is Leo.

Why do we need Memory ?

LLMs are generally stateless between individual calls.
stateless --> Meaning --> They won't remember any past conversation.
Memory solves this by maintaining relevant conversation information and providing it to the model with next request.

Stateless Vs State-full LLM

Stateless means the LLM doesn't remember previous request automatically. Each request treated independently.
State-full means the application maintains information about previous interactions and provide relevant state to the LLM.

Key in Memory

Stateless to Statefull is the key idea behind LangChain Memory.

Session ID

Statefull applications usually need to identify whose conversation is being maintained.

Overall

LLM usually Stateless.
Application + Memory/State --> Statefull Experience
A Stateless LLM treats each API request independently and doesn't automatically retain previous conversation context.
A Statefull LLM application maintains state such as conversation history and provides that state with subsequent requests , allowing the system to maintain context across interactoins.

Conversation Buffer Memory in LangChain

Simplest type of memory.
It stores complete conversation history and sends that history along with every new question to the LLM.

Without Memory:

Because the second request doesn't automatically contain the first conversation.

With Conversation Buffer Memory:

Memory stores the previous messages.

Main Disadvantage :

Suppose the conversation becomes very long.
Every new request may require a large amount of history to be sent to the model.
This will lead to :

Higher Token Usage
Higher Cost
Slower Response
Context Window Limitations

Conversation Buffer Window in LangChain

Keeps only the most recent part of the conversation instead of the complete conversation.
The word WINDOW means we maintain a moving window of recent messages.

1. Buffer Memory --> Remembers everything 
2. Buffer Window Memory --> Remembers only the recent messages

Why we need window memory ?

Conversation Token Buffer Memory in LangChain

Its a token-based memory approach.
Instead of saying: " Keep the last 4 messages "
It says : Keep the conversation history only up to a certain number of tokens.
This is useful because LLMs process tokens, not messages.

Conversation Summary Memory in LangChain

It keeps a summary of the previous conversation instead of sending the entire conversation to the LLM every time.
Its is very useful when the conversation becomes long.
Conversation:

Summary:

LangChain Agent

Agent places important role when it comes to Generative AI.
Its an application where an LLM can decide what action to take, use available tools, observe the result and continue until it can answer the user's request.
Agent can decide which tool needs to be used.
Example:

Easy to REMEMBER:

Chain --> You can decide the steps
Agent --> LLM decides the steps.

Agent Architecture

The Four Main Components

LangChain Callbacks

Callbacks are the mechanisms that allow you to monitor and respond to events that happen during a LangChain execution.
Callbacks let us observe what is happening inside an LLM application.