API Reference

LangChain's API is intentionally small at the seams: most of the symbols below are all you'll import from langchain-core, and the rest come from integration packages that implement the same interface.

Core runnables (langchain_core.runnables)

• Runnable — the base interface (invoke/ainvoke/batch/abatch/stream/astream).
• RunnableSequence — produced by a | b.
• RunnableLambda — wrap a plain function.
• RunnablePassthrough — pass input through unchanged (attach context).
• RunnableParallel / RunnableMap — fan-out over the same input.
• RunnableBranch — conditional routing.

Messages and models

• BaseMessage, HumanMessage, SystemMessage, AIMessage, ToolMessage.
• BaseChatModel — the interface; bind_tools(), invoke(messages), stream().
• ChatOpenAI (from langchain-openai), ChatAnthropic (from langchain-anthropic) — concrete chat models.

Prompts

• ChatPromptTemplate.from_messages(...) — the main prompt builder.
• PromptTemplate.from_template("...") — simple string templates.
• MessagesPlaceholder("history") — inject a dynamic message list (memory, scratchpad).

Output parsers

• StrOutputParser — raw string.
• JsonOutputParser / PydanticOutputParser — structured output.

Retrieval

• BaseRetriever — the query→documents interface.
• RecursiveCharacterTextSplitter (from langchain-text-splitters) — chunking.
• OpenAIEmbeddings (from langchain-openai) — embeddings.
• FAISS, InMemoryVectorStore, pgvector, etc. — vector stores with .as_retriever().

Tools and agents

• @tool — decorator turning a typed function into a BaseTool.
• create_tool_calling_agent(model, tools, prompt) — the modern agent.
• AgentExecutor(agent=..., tools=...) — Runnable-style agent loop with max_iterations.
• StateGraph (from langgraph) — graph-based agents with compile() and persistence.

Chain helpers (langchain.chains)

• create_retrieval_chain(retriever, combine_docs_chain) — the RAG wrapper.
• create_stuff_documents_chain(llm, prompt) — "stuff all docs into the prompt" step.

Edge cases

• Import paths changed across LangChain versions (e.g. things moving from langchain to langchain-community to their own package) — pin versions and verify imports at startup, not at first use.
• bind_tools output is provider-specific JSON; always parse tool calls with the model class's own .tool_calls accessor, not by eyeballing the raw message.
• Streaming a full chain streams each Runnable's output — the intermediate prompt/retrieval steps are not tokens, so UI code should subscribe to the final model step (or LangGraph's intermediate streaming) for clean UX.