langchain-components（LangChain 组件）

LangChain Components

Complete reference for the LangChain ecosystem — models, agents, tools, retrieval, memory, middleware, streaming, multi-agent orchestration, LangGraph workflows, Deep Agents, and provider integrations for Python 3.10+.

Component Index

Models & Output

• Models — Chat models, tool calling, multimodal inputs, caching, rate limiting, custom models reference
• Messages — Message types (Human, AI, System, Tool), message operations, serialization, OpenAI format conversion reference

Agents

• Agents — create_agent, tools, structured output, guardrails, human-in-the-loop, context engineering reference
• Multi-Agent — Subagents, handoffs, skills, router, custom workflows, pattern selection reference

Tools & MCP

• Tools — Tool creation (@tool decorator, ToolNode), InjectedState, MCP integration, error handling reference

Retrieval & RAG

• Retrieval — Document loaders, text splitters, embeddings, vector stores, agentic RAG, semantic search reference

Memory

• Memory — Short-term (checkpointers, message trimming, summarization), long-term (store abstraction, namespaces) reference

Middleware & Streaming

• Middleware — 16 built-in middleware, custom middleware (decorator, class, wrap-style), execution order reference
• Streaming — Stream modes (updates, messages, custom), token streaming, useStream React hook reference

Runtime & Architecture

• Runtime — Dependency injection, context schemas, ToolRuntime, component architecture (5 layers) reference

Testing & Deployment

• Testing — Unit testing (GenericFakeChatModel), integration testing (AgentEvals), LangSmith observability reference

LangGraph

• LangGraph Core — Graph API, Functional API, workflows vs agents, state management, quickstart reference
• LangGraph State — Memory, persistence, durable execution, interrupts, checkpointers reference
• LangGraph Advanced — Subgraphs, time-travel, streaming, Graph API usage, Functional API usage reference

Deep Agents

• Deep Agents — Harness framework, models, subagents, skills, sandboxes, human-in-the-loop, long-term memory reference

Integrations

• Integrations — Chat models, document loaders, retrievers, embeddings, vector stores, tools, stores, splitters reference
• Providers — OpenAI, Anthropic, Google, AWS, Ollama setup and configuration reference

Quick Patterns

Create an Agent with Tools

from langchain.chat_models import init_chat_model
from langgraph.prebuilt import create_agent

model = init_chat_model("anthropic:claude-sonnet-4-20250514")

def get_weather(city: str) -> str:
    """Get weather for a city."""
    return f"Sunny, 72F in {city}"

agent = create_agent(model, [get_weather])
response = agent.invoke(
    {"messages": [{"role": "user", "content": "What's the weather in SF?"}]}
)

Structured Output

from pydantic import BaseModel

class SearchQuery(BaseModel):
    query: str
    year: int

structured_model = model.with_structured_output(SearchQuery)
result = structured_model.invoke("Who won the World Cup in 2022?")

RAG with Retrieval

from langchain_community.document_loaders import WebBaseLoader
from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain_openai import OpenAIEmbeddings
from langchain_core.vectorstores import InMemoryVectorStore

docs = WebBaseLoader("https://example.com").load()
chunks = RecursiveCharacterTextSplitter(chunk_size=1000).split_documents(docs)
vector_store = InMemoryVectorStore.from_documents(chunks, OpenAIEmbeddings())
retriever_tool = vector_store.as_retriever()

Multi-Agent Handoffs

from langgraph.prebuilt import create_agent

billing_agent = create_agent(model, [lookup_billing], name="billing")
tech_agent = create_agent(model, [check_status], name="tech_support")
supervisor = create_agent(
    model,
    [billing_agent, tech_agent],
    prompt="Route to the appropriate specialist."
)

LangGraph Workflow

from langgraph.graph import StateGraph, START, END

graph = StateGraph(dict)
graph.add_node("process", process_fn)
graph.add_node("review", review_fn)
graph.add_edge(START, "process")
graph.add_edge("process", "review")
graph.add_edge("review", END)
app = graph.compile()

Streaming

for chunk in agent.stream(
    {"messages": [{"role": "user", "content": "Hello"}]},
    stream_mode="messages"
):
    print(chunk)

Best Practices

• Use init_chat_model() for provider-agnostic model initialization
• Prefer create_agent over building custom agent loops
• Use LangGraph for complex workflows requiring state, persistence, or human-in-the-loop
• Apply middleware for cross-cutting concerns (guardrails, rate limiting, PII detection)
• Use checkpointers for conversation persistence and short-term memory
• Use the Store abstraction for long-term memory across conversations
• Choose the right multi-agent pattern: handoffs for specialization, routers for classification, subagents for parallel work
• Use with_structured_output() for type-safe LLM responses
• Prefer agentic RAG (tool-based retrieval) over chain-based RAG for flexibility
• Use stream_mode="messages" for token-level streaming to frontends