Primitives

The cross-cutting abstractions that every other package builds on. They all live in libs/core/langchain_core/.

Primitive	Purpose	Page
Runnables / LCEL	Universal invoke/batch/stream protocol; the | operator	runnables
Messages and content blocks	Conversation turn types and the typed-content taxonomy	messages
Language models	BaseChatModel, BaseLLM, BaseChatModelV1	language-models
Tools	BaseTool, StructuredTool, the @tool decorator	tools
Prompts	PromptTemplate, ChatPromptTemplate, few-shot variants	prompts
Output parsers	Turn model output into typed values	output-parsers
Callbacks and tracers	The observability backbone	callbacks-and-tracers

These are the symbols that show up in almost every example, every integration, and every downstream package. Understanding them is the prerequisite for reading the rest of the codebase.

Why "primitives"?

Each item on this page is something users build with, but never look inside. A Runnable is composable; you don't need to know how RunnableSequence.invoke walks its inner steps to use prompt | model | parser. A BaseChatModel is a contract; you don't need to know how ChatOpenAI translates content blocks to use it via init_chat_model.

The pages in this section explain the contracts — what each primitive guarantees, what implementations exist, and how downstream code consumes them.

Layering

graph TD
    Runnable["Runnable / LCEL"]
    Msg[Messages + Content Blocks]
    LM[Language Models]
    Tool[Tools]
    Prompt[Prompts]
    Out[Output Parsers]
    CB[Callbacks / Tracers]
    Doc[Documents / VectorStores]

    Prompt --> Runnable
    LM --> Runnable
    Tool --> Runnable
    Out --> Runnable
    Doc --> Runnable
    LM --> Msg
    Tool --> Msg
    Runnable --> CB
    LM --> CB
    Tool --> CB

Runnable sits at the bottom — almost everything else inherits from it. Messages flow through Language models and Tools. Callbacks observe everything.

Where to go next

• The overview/architecture page shows how these primitives compose into agents and chains.
• The packages/core page lists every module in libs/core/langchain_core/.
• The features section covers the higher-level patterns built on top: agents, middleware, structured output.