TopolearnLangChain is split into many packages; knowing which one does what avoids import errors and outdated tutorials.
LangChain is a family of packages rather than one library:
- langchain-core: the base abstractions (messages, chat model and embedding interfaces, prompts, runnables, tools, documents, vector store interface). Everything else depends on it.
- langchain: the high-level entry point. In 1.x it is deliberately small:
init_chat_model,create_agentwith middleware, and re-exports of common core pieces. - Provider packages such as
langchain-openai,langchain-anthropic,langchain-google-genai,langchain-ollama: concrete model and embedding classes. Install only the providers you use. - langchain-text-splitters and langchain-community: splitters, document loaders and many community integrations.
- LangGraph: a lower-level library for stateful, multi-step agent workflows with persistence, streaming and human-in-the-loop. LangChain agents run on it.
- langchain-classic: legacy chains and retrievers from 0.x (such as
RetrievalQA) kept for migration. - LangSmith: a hosted platform for tracing, evaluation and deployment, configured with environment variables.
A lot of material online targets 0.1 to 0.3 and uses APIs like LLMChain, initialize_agent or AgentExecutor, which are legacy now. In 1.x the building blocks are chat models, messages, LCEL runnables, tools and create_agent/LangGraph.
API keys are read from environment variables (OPENAI_API_KEY, ANTHROPIC_API_KEY...). Keep them out of code.
Example
# pip install -U langchain langchain-openai langgraph
import importlib.metadata as md
for pkg in ["langchain", "langchain-core", "langgraph", "langchain-openai"]:
try:
print(f"{pkg:18} {md.version(pkg)}")
except md.PackageNotFoundError:
print(f"{pkg:18} not installed")
# langchain 1.x
# langchain-core 1.x
# langgraph 1.x
# langchain-openai 1.x
from langchain_core.messages import HumanMessage # core abstractions
from langchain.chat_models import init_chat_model # high-level helpers
from langchain.agents import create_agent # agents (built on LangGraph)Remember
- langchain-core holds abstractions; provider packages hold model integrations
- LangGraph powers agents and stateful workflows
- Pre-1.0 APIs like LLMChain and AgentExecutor are legacy (langchain-classic)
Watch out for
- Following a 0.1-era tutorial that imports from langchain.llms or uses initialize_agent, which no longer exists in 1.x.