- API calls often involve extended execution times, delivering outputs progressively as they're generated
- Unlike structured inputs with defined parameters (e.g., JSON), they process unstructured, free-form natural language, comprehending its nuances
- Results are nondeterministic - identical inputs may yield different outputs
LangChain emerges as a leading framework for creating LLM-driven applications, addressing these challenges and providing extensive integrations with proprietary model providers (OpenAI, Anthropic, Google), open-source alternatives, and complementary components like vector stores.
This guide explores the fundamentals of building applications with LLMs using LangChain's Python library. Only basic Python knowledge is required - no machine learning background necessary!
What You'll Learn:
- Initial project configuration
- Working with chat models and core LangChain components
- Constructing chains using LangChain Expression Language
- Implementing real-time streaming responses
- Providing context to guide model outputs (basic RAG concepts)
- Debugging and tracing chain internals
Let's dive in!
Project Configuration
We recommend using Jupyter notebooks for this tutorial's code, offering an interactive, clear environmnet. Follow these setup instructions for local installation, or use Google Colab for a browser-based experience.
First, select your preferred chat model. If you've used ChatGPT-like intefraces, you'll find chat models familiar - they accept messages as input and return messages as output. The distinction is we'll manage this programmatically.
This guide defaults to Anthropic's Claude 3 chat model, but LangChain offers numerous other integrations, including OpenAI's GPT-4.
pip install langchain_core langchain_anthropic
In Jupyter notebooks, prepend % to pip: %pip install langchain\_core langchain\_anthropic.
You'll also need an Anthropic API key from their console. Set it as the ANTHROPIC\_API\_KEY environment variable:
export ANTHROPIC_API_KEY="..."
Alternatively, pass the key directly to the model constructor.
Getting Started
Initialize your model like this:
from langchain_anthropic import ChatAnthropic
conversation_model = ChatAnthropic(
model="claude-3-sonnet-20240229",
temperature=0
)
# For explicit key passing:
# conversation_model = ChatAnthropic(
# model="claude-3-sonnet-20240229",
# temperature=0,
# api_key="YOUR_ANTHROPIC_API_KEY"
# )
The model parameter matches one of Anthropic's supported models. Currently, Claude 3 Sonnet offers an optimal balance of speed, cost, and reasoning capabilities.
temperature controls response randomness. We'll use 0 for consistency, but experiment with higher values for creative applications.
Now let's test it:
conversation_model.invoke("Share a programming joke!")
Output:
AIMessage(content="Here's a programming joke for you:\n\nWhy do programmers prefer dark mode?\nBecause light attracts bugs!")
Notice the output is an AIMessage object. Chat models use message objects for both input and output.
Note: The previous example accepted a plain string because LangChain provides convenient shorthand conversion. A single string automatically becomes a HumanMessage array.
LangChain also includes text completion LLMs with string inputs/outputs. However, chat-focused models have surpassed them in popularity - GPT-4 and Claude 3 are chat models.
For clarity, let's explicitly pass a message array:
from langchain_core.messages import HumanMessage
conversation_model.invoke([
HumanMessage("Tell me a joke about bears!")
])
Similar output:
AIMessage(content="Here's a bear joke for you:\n\nWhy did the bear bring a briefcase to work?\nHe was a business bear!")