Core Framework Structure

LangChain consists of three primary packages:

• LangChain Core: Contains fundamental data structures and the LangChain Expression Language (LCEL)
• LangChain Community: Open-source integrations and community-contributed components
• LangChain Applications: High-level implementation logic for building applications

Key Terminology

LLM (Large Language Model)

Base models focused on text completion and generation tasks.

RAG (Retrieval Augmented Generation)

Combines retrieval models with generative models to enhance output relevance and accuracy. This approach reduces hallucinations and improves integration with external knowledge sources.

LCEL (LangChain Expression Language)

A domain-specific language using pipe operators (|) to connect Runnable components into processing chains:

processing_chain = prompt_template | language_model | output_parser

Core Components

Model Wrappers

Abstracts API differences between various LLM providers, simplifying integration.

Text Completion Models (LLM):

from langchain.llms import OpenAI

text_model = OpenAI(model="text-davinci-003")
response = text_model("Describe the highest mountain in China")
print(response)

Chat Models:

from langchain.chat_models import ChatOpenAI
from langchain.prompts import (
    SystemMessagePromptTemplate,
    HumanMessagePromptTemplate,
    ChatPromptTemplate
)

chat_model = ChatOpenAI()

system_template = SystemMessagePromptTemplate.from_template(
    "You are an AI assistant"
)
user_template = HumanMessagePromptTemplate.from_template("{query}")

chat_prompt = ChatPromptTemplate.from_messages(
    [system_template, user_template]
)

response = chat_model(chat_prompt.format_prompt(query="Tell me about Mount Everest").to_messages())
print(response.content)

Prompt Templates

Structured templates for generating consistant model inputs:

# Similar to Python's string formatting
from langchain.prompts import PromptTemplate

template = PromptTemplate.from_template(
    "Tell me about {subject} in {language} language"
)
formatted = template.format(subject="quantum physics", language="simple")