LangChain Overview

by Carson Kempf

LangChain

• LangChain is a framework for LLM applications

Introduction

Installation & Setup

1. Install package

   pip install -qU "langchain[deepseek]" 
   

2. Get Environment Variables and Chat Model

import getpass
import os

if not os.environ.get("OPENAI_API_KEY"):
  os.environ["OPENAI_API_KEY"] = getpass.getpass("Enter API key for OpenAI: ")

from langchain.chat_models import init_chat_model

model = init_chat_model("gpt-4o-mini", model_provider="openai")

1. First API Call

   model.invoke("Hello, world!")
   

Architecture

Some Open Source Libraries

langchain-core

• Basic chat model features

  langchain-deepseek
  

• Integration package specifically for the deepseek chat models

  langchain
  

• Chains, agents, and retrieval strategies that orchestrate an application’s actual, logically implemented features

  langgraph
  

• Framework for orchestration
• Represents an application’s LLM steps as edges and nodes in a graph

Tutorials

Simple LLM with Chat Model and Prompt Templates

1. Installation

   pip install langchain
   

Note: Use LangSmith!

• Used to construct more complex applications
• Can inspect agents and chains more easily
• LangSmith Environment Variables

export LANGSMITH_TRACING="true"
export LANGSMITH_API_KEY="..."
export LANGSMITH_PROJECT="default" # or any other project name

1. Simple Call to a Chat Model

from langchain_core.messages import HumanMessage, SystemMessage

messages = [
    SystemMessage("Translate the following from English into Italian"),
    HumanMessage("hi!"),
]

model.invoke(messages)

• A ChatModel (model.invoke) is a LangChain runnable
• LangChain Runnables
  • Expose an interface to interact with a chat model
• Example Langchain Message

  AIMessage(content='Ciao!', additional_kwargs={'refusal': None}, response_metadata={'token_usage': {'completion_tokens': 3, 'prompt_tokens': 20, 'total_tokens': 23, 'completion_tokens_details': {'accepted_prediction_tokens': 0, 'audio_tokens': 0, 'reasoning_tokens': 0, 'rejected_prediction_tokens': 0}, 'prompt_tokens_details': {'audio_tokens': 0, 'cached_tokens': 0}}, 'model_name': 'gpt-4o-mini-2024-07-18', 'system_fingerprint': 'fp_0705bf87c0', 'finish_reason': 'stop', 'logprobs': None}, id='run-32654a56-627c-40e1-a141-ad9350bbfd3e-0', usage_metadata={'input_tokens': 20, 'output_tokens': 3, 'total_tokens': 23, 'input_token_details': {'audio': 0, 'cache_read': 0}, 'output_token_details': {'audio': 0, 'reasoning': 0}})
  

Message Objects

• ChatModels receive message objects as inputs
• They generate message objects as outputs in return
• Message objects store a role for the chat
• Message objects store other useful metadata like token usage counts and tool calls
• OpenAI Format

model.invoke("Hello")

model.invoke([{"role": "user", "content": "Hello"}])

model.invoke([HumanMessage("Hello")])

Streaming

• Remember: Chat Models are Runnables
• Runnables expose interfaces from the API to the developer
• Runnables interface includes
  • Async mode of invocation
  • Streaming mode of invocation
• This way, we can stream individual tokens from a chat model
• Streaming tokens as they arrive instead of once they are all compiled and sent at once

• Streaming Usage

  for token in model.stream(messages):
      print(token.content, end="|")
  

Prompt Templates

• Transformations
  • Often transformations are made on the user input before being sent to the AI model
  • Including
    • Adding a system message or metadata
    • Formatting a template with user input
• Example ``` from langchain_core.prompts import ChatPromptTemplate

system_template = “Translate the following from English into {language}”

prompt_template = ChatPromptTemplate.from_messages( [(“system”, system_template), (“user”, “{text}”)] )

language

* The language to translate text into

text

* The text to be translated

* Note:
  * ChatPromptTemplate supports multiple message roles
  * In our example we formatted the language variable as a system message
  * We formatted the user text into a user message

* How to Invoke a Template with Correct Parameters

prompt = prompt_template.invoke({“language”: “Italian”, “text”: “hi!”})

prompt

* What  **prompt**  Returns:

ChatPromptValue(messages=[SystemMessage(content=’Translate the following from English into Italian’, additional_kwargs={}, response_metadata={}), HumanMessage(content=’hi!’, additional_kwargs={}, response_metadata={})])

* What **prompt.to_messages()** Returns:

[SystemMessage(content=’Translate the following from English into Italian’, additional_kwargs={}, response_metadata={}), HumanMessage(content=’hi!’, additional_kwargs={}, response_metadata={})]

* How to Actually Get a Response From our Invocation:

response = model.invoke(prompt) print(response.content) ```

Topics To Explore Next

• Conceptual Guide
• Tools / Function Calling
  • For some reason, a separate tool calling doc exists
• Agents / Calling External APIs
  • LangGraph is recommended for constructing agents