The most-used framework for composing LLM calls, tools, and memory into agents and pipelines — batteries-included but with real opinions.
LangChain is the Swiss Army knife for building LLM-powered applications. It provides abstractions for prompts, models, embeddings, vector stores, document loaders, text splitters, memory, tools, and agents — all with integrations to virtually every model provider, vector database, and data source you can name. The framework has matured considerably since 2023. The current version is built on LangGraph, a graph-based runtime for composing stateful, multi-step agents. You can build simple chains in a few lines or complex agentic workflows with cycles, human-in-the-loop, and persistence. LangSmith is the commercial observability and evaluation layer. The honest trade-off: LangChain is opinionated and big. The abstractions are useful when you need them and overhead when you don't. For a 10-line prompt to OpenAI, plain `fetch` is simpler. For a 50-node agent with branching, retries, and observability, LangGraph is genuinely productive.
Graph-based composition of multi-step agents with cycles, persistence, and human-in-the-loop.
OpenAI, Anthropic, Google, Mistral, Cohere, local Ollama, vLLM — all behind one interface.
Pinecone, Weaviate, Chroma, pgvector, Qdrant, LanceDB — swap with one config change.
Tracing, evaluation, dataset management, and prompt versioning in one tool.
PDFs, Notion, Slack, GitHub, Confluence, S3 — load from anywhere with the same interface.
Login to comment
Migrated from a hand-rolled RAG pipeline last quarter. The vector store abstraction alone was worth it — we swap from FAISS to pgvector in one config line.
If you're doing a 10-line prompt to OpenAI, skip the framework. If you're doing a 50-node agent with cycles and persistence, use LangGraph. The middle ground is where teams over-engineer with LangChain.
Migrated from a hand-rolled RAG pipeline last quarter. The vector store abstraction alone was worth it — we swap from FAISS to pgvector in one config line.
If you're doing a 10-line prompt to OpenAI, skip the framework. If you're doing a 50-node agent with cycles and persistence, use LangGraph. The middle ground is where teams over-engineer with LangChain.
Migrated from a hand-rolled RAG pipeline last quarter. The vector store abstraction alone was worth it — we swap from FAISS to pgvector in one config line.
If you're doing a 10-line prompt to OpenAI, skip the framework. If you're doing a 50-node agent with cycles and persistence, use LangGraph. The middle ground is where teams over-engineer with LangChain.