LangChain

LangChain is a framework for building LLM applications. It provides abstractions for prompts, chat models, tools, memory, and composition. This section teaches LangChain from first principles, building concepts slowly before putting them together.

---

Learning path

Read these in order. Each file builds on the previous one.

#	File	What you learn
1	01-mental-model	What LangChain solves, the four core concepts (messages, models, tools, runnables), chains vs agents
2	02-messages	The five message types (SystemMessage, HumanMessage, AIMessage, ToolMessage, AIMessageChunk), the alternation invariant, RemoveMessage
3	03-chat-models	ChatOpenAI, bind_tools, with_structured_output, with_retry, with_fallbacks, provider kwargs
4	04-tools	@tool decorator, schema from type annotations, docstrings as API contract, async tools, error handling, InjectedToolArg
5	05-prompts	ChatPromptTemplate, the four slot types, MessagesPlaceholder, partial, few-shot, RAG context placement
6	06-runnables-lcel	The Runnable interface, the `
7	07-memory-callbacks	set_llm_cache, BaseCallbackHandler, the hook lifecycle, LangSmith tracing, why legacy memory classes are deprecated
8	08-langgraph-intro	StateGraph, nodes, edges, ToolNode, checkpointers (MemorySaver, SqliteSaver), Command, interrupts
9	09-streaming	stream/astream, astream_events (full lifecycle API), building a streaming chat UI, batch/abatch
10	10-testing	FakeListChatModel, FakeMessagesListChatModel, testing the graph, no-network discipline, patch_langchain_environment

---

Prerequisites

• Python 3.11+
• An OpenAI-compatible API endpoint (OpenAI, Anthropic via LiteLLM, AWS Bedrock via LiteLLM, or local Ollama)

# pyproject.toml
dependencies = [
    "langchain-core>=0.3",
    "langchain-openai>=0.3",
    "langgraph>=0.2",
]

---

The progression

01-mental-model        ← start here
       ↓
02-messages            ← the atom everything uses
       ↓
03-chat-models         ← call a model, get back AIMessage
       ↓
04-tools               ← model can call functions
       ↓
05-prompts             ← template the messages list
       ↓
06-runnables-lcel      ← compose with |
       ↓
07-memory-callbacks    ← cache, observe, trace
       ↓
08-langgraph-intro     ← agent loop (the goal)
       ↓
09-streaming           ← token-by-token output
       ↓
10-testing             ← test without paying

By the end, you can build a full agent: a graph with a call_model node, a ToolNode, checkpointing for conversation persistence, and a streaming UI.

---

What this section is NOT

This is not the LangChain documentation. It is a learner’s guide. Every concept is introduced from scratch, with simple examples that don’t reference a specific project. The focus is on building mental models, not API signatures.

For the authoritative docs, see python.langchain.com.

---

See also

• LangGraph — the agent framework built on LangChain (comes after this section)
• llm-foundations — the machine learning concepts underneath LLMs