LangChain Essentials Cheat Sheet

LangChain framework for LLM applications — chains, agents, tools, memory, retrieval, and the LangChain Expression Language (LCEL) for composable pipelines.

Last Updated: May 1, 2025

Core Components

Item	Description
`LLMs/Chat Models`	Interface to any model — OpenAI, Anthropic, HuggingFace, local via Ollama
`Prompt Templates`	Parameterized prompts with input variables — reuse and compose
`Chains`	Sequence of calls — prompt → LLM → parser → output. LCEL: prompt \| llm \| parser
`Retrievers`	Fetch relevant documents from vector stores — RAG backbone
`Agents`	LLM decides which tool to use and when — ReAct, OpenAI functions, structured chat
`Memory`	Persist conversation state — buffer, summary, vector-store backed

LCEL (Expression Language)

chain = prompt | llm | StrOutputParser()

Basic chain: template → generate → parse string

chain = prompt | llm | JsonOutputParser()

JSON output chain — LLM returns structured data

chain = {'context': retriever, 'question': RunnablePassthrough()} | prompt | llm

RAG chain — retrieves + passes context

chain.with_config(run_name='MyChain')

Name for tracing in LangSmith

chain.stream(input)

Stream response token-by-token

chain.batch([input1, input2])

Process multiple inputs in parallel

Retrieval (RAG) Patterns

Item	Description
`Document Loaders`	Load from 100+ sources: PDF, web, database, Notion, GitHub
`Text Splitters`	RecursiveCharacterTextSplitter — chunk by paragraph/sentence with overlap
`Embeddings`	OpenAI, Cohere, HuggingFace — convert text to vectors
`Vector Stores`	Chroma (local), Pinecone (cloud), Weaviate, Qdrant, FAISS
`RetrievalQA`	Question-answering over documents with source attribution
`ConversationalRetrieval`	RAG with chat history — rephrase question using history then retrieve

Ecosystem

Item	Description
`LangSmith`	Observability platform — traces, evals, datasets, prompt hub
`LangServe`	Deploy chains as REST APIs — FastAPI + chain = production endpoint
`LangGraph`	Stateful multi-actor applications — graph-based agent workflows
`LangChain Hub`	Community prompt templates — import, fork, customize

Pro Tip: Use LCEL (LangChain Expression Language) for all new chains — it's more composable, supports streaming, and is easier to debug than the legacy Chain classes.