LangChain

Python/JS framework for chains, agents, and RAG. Batteries-included but heavyweight; many graduate to LangGraph or DIY.

Install via pip: pip install langchain langchain-community langchain-openai. Requires Python 3.10+. LangChain is the oldest and most comprehensive framework for building LLM-powered applications — chains, agents, RAG pipelines, and tool use. Basic chat: from langchain_openai import ChatOpenAI; llm = ChatOpenAI(model="gpt-4o"); response = llm.invoke("Hello"). For a simple RAG chain: pip install langchain chromadb then from langchain_community.document_loaders import TextLoader; from langchain_text_splitters import RecursiveCharacterTextSplitter; from langchain_openai import OpenAIEmbeddings; from langchain_chroma import Chroma; from langchain.chains import create_retrieval_chain; from langchain.chains.combine_documents import create_stuff_documents_chain; from langchain_core.prompts import ChatPromptTemplate. LangChain uses a modular provider architecture: langchain-openai, langchain-anthropic, langchain-ollama, etc. Support for local models: pip install langchain-ollama then ChatOllama(model="llama3.2"). LangChain's LCEL (LangChain Expression Language) composes chains with | pipe syntax. First run: ~2 minutes for package install, + model API latency. Verify: run the chat example above and confirm a response. Time-to-first-prototype: ~5 minutes.

Best for: complex LLM application orchestration with many integrated services (databases, APIs, vector stores, document loaders), enterprise LLM applications where integration breadth matters more than abstraction depth, teams that value community size and availability of answers, RAG pipelines needing diverse document format support, prototyping with pre-built chains before production-engineering. Not suited for: simple LLM tasks where direct API calls suffice, production systems requiring fine-grained control and predictable behavior (LangChain's abstractions are leaky at scale), teams that have been burned by LangChain version migration churn (v0.1→v0.2→v1.0), applications prioritizing latency where framework overhead matters, users preferring minimal-dependency stacks.

Use LangChain when you need the broadest ecosystem for LLM application development — 100+ integrations (databases, vector stores, document loaders, tools), 1000+ community extensions, and the largest community for answers. It's the default framework for many enterprise LLM apps. Switch to LlamaIndex when your primary use case is RAG/data ingestion — LlamaIndex's document parsing and indexing pipeline is more polished. Use Haystack for a more opinionated, pipeline-based alternative. Use direct provider SDKs (OpenAI, Anthropic) when you need maximum control and minimal abstraction — LangChain's abstractions can hide behavior that matters for production. Use DSPy when you want programmatic prompt optimization rather than chain orchestration. LangChain's strength: breadth — if an integration exists, LangChain likely has it. Its weakness: abstraction depth obscures what's happening under the hood, leading to debugging difficulty and surprise behavior in production. The v0.1→v0.2→v1.0 migration path has been painful for many teams.

Problem: ImportError: cannot import name 'X' from 'langchain'. Fix: LangChain split into multiple packages in v0.2+: langchain-core, langchain-community, langchain-openai, etc. The legacy langchain package still exists but imports should use the specific sub-packages. Update import paths: from langchain.chains → from langchain.chains, from langchain.llms → from langchain_openai or equivalent provider package. Check the LangChain v0.2 migration guide for exact mappings. Problem: Chain execution produces unexpected results or silent errors. Fix: LangChain's chain composition hides intermediate steps. Enable verbose mode: chain = chain.with_config({"verbose": True}) or use LangSmith tracing (LANGCHAIN_TRACING_V2=true) to see every step. LCEL chains with many pipe stages can swallow exceptions — wrap in try/except and check RunnableSequence intermediate outputs. Problem: RAG retrieval returns irrelevant documents despite vector DB having correct data. Fix: The default retriever may use a different embedding model than what indexed the documents. Ensure embeddings in the retriever matches the vector store's indexing embedding model. Mismatched embedding dimensions are the most common silent RAG failure — the retriever returns cosine-distance-sorted noise.