LangChain course: build real LLM apps in Python

Explain what LangChain adds on top of a plain model call and set up a safe project.

• What an LLM app is: model, prompt, tools, data and your code
• What LangChain gives you and when a plain API call is enough
• Model providers (OpenAI, Google Gemini, Anthropic Claude, local models) behind one interface
• Project setup: virtual environment, environment variables and keeping API keys out of Git
• Reading the official documentation, and why versions matter

PythonLangChainVS CodeGit

Practice: A first script that sends one question to a model and prints the answer

Write prompt templates and chain steps so the same app works with different inputs.

• Chat messages: system, user and assistant roles
• Prompt templates with variables and examples
• Chaining steps: prompt, model, output parser
• Temperature and other settings, and what they change
• Streaming a reply as it is written

LangChainLLM provider SDK

Practice: A reusable email or summary writer that takes a topic and a tone

Make a model return data your code can trust, and let it call your functions.

• Structured output with typed schemas (for example Pydantic models)
• Validating output and handling a bad reply
• Tool or function calling: describing a tool so a model can choose it
• Writing safe tools: input checks, timeouts and clear error messages
• Calling an outside API from a tool, such as weather or currency

PydanticLangChain tools

Practice: A resume or invoice reader that returns clean JSON · A small assistant with two tools

Build a question-and-answer app that answers from your own files and shows its sources.

• Why models need outside data: stale knowledge and private documents
• Loading PDFs, text and web pages, and splitting them into chunks
• Embeddings in simple words: turning meaning into numbers
• Vector stores and similarity search
• Retriever plus prompt plus model, with sources shown in the answer
• Common RAG failures: bad chunks, missing context, made-up answers

EmbeddingsA vector store (open-source or local)LangChain

Practice: Ask-my-notes: question answering over a set of PDFs with page references

Build a chatbot that remembers the conversation and stays within the model's limits.

• Why models forget: every call starts fresh
• Passing message history, and trimming or summarising long chats
• Separate conversations for separate users
• Short-term memory versus saved memory such as user preferences
• Privacy: what you should not store

LangChainSQLite

Practice: A study-helper chatbot that remembers the topic and the student's level

Build a simple agent that decides which tool to use and knows when to stop.

• What an agent is: a model that loops, calls tools and reads the results
• Choosing tools and writing good tool descriptions
• Limits: maximum steps, timeouts and what happens when a tool fails
• When a fixed chain is better than an agent
• Where LangGraph fits when you need more control over the flow

LangChain agents

Practice: A research helper that searches, reads and summarises with a step limit

See what your app is doing, measure its quality and keep cost and risk under control.

• Tracing each step of a run with LangSmith or a similar tool
• Building a small test set of questions with expected answers
• Simple evaluations: exact checks, rule checks and model-as-judge, and the limits of each
• Token counts and cost: caching, shorter prompts and smaller models for easy steps
• Prompt injection, private data and sensible guardrails

LangSmithpytest

Practice: Add tracing and a 20-question test set to your RAG app and report what fails

Put your app behind an API, deploy it and present it with a clear README.

• Wrapping your chain or agent in a small web API
• Configuration, secrets and rate limits
• Logging, error handling and timeouts
• A simple front end or a chat page that calls your API
• Deploying on a free or low-cost host and writing the README and demo

FastAPIRender or RailwayGitGitHub

Practice: Capstone: a deployed app with an API, tracing and a public GitHub repository