Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
The user only needs to call an Anthropic SDK or OpenAI SDK directly without chains / retrievers / tools — use claude-api or openai-sdk instead
The task is pure graph orchestration with custom state machines, no LangChain primitives — use the langgraph skill (LangGraph is a separate package; create_agent from langchain.agents is built on top of it)
Serving the chain behind HTTP — use fastapi for the HTTP layer and call LangChain inside the handler
The user wants raw prompt-engineering tips with no orchestration code — use the model-vendor skill
The codebase is locked to deprecated AgentExecutor and the user is not refactoring — skill bias is toward v1 idioms, not maintaining v0 hacks
Purpose
LangChain Python 1.x is a composable framework for building LLM applications. The core abstraction is Runnable — anything with .invoke(), .ainvoke(), .stream(), .astream(), .batch() — and the | pipe operator that wires Runnables into chains (LCEL). On top of this, LangChain ships chat-model adapters (langchain-anthropic, langchain-openai, ...), prompt templates, output parsers, tool decorators, retriever and vector-store interfaces, plus first-class streaming and LangSmith observability.
v1 is a meaningful break from v0. The Chain class hierarchy is retired in favor of plain LCEL composition. The old AgentExecutor is deprecated; agents now come from langchain.agents.create_agent (built on LangGraph). init_chat_model is the unified provider-agnostic entry point. Memory has moved into LangGraph checkpointers — RunnableWithMessageHistory still exists but new code typically uses a LangGraph state graph with a checkpointer for multi-turn state. This skill codifies the v1 idioms and flags v0 anti-patterns that still appear in stale tutorials.
Capabilities
Provider-agnostic model init
Use init_chat_model("provider:model-id") instead of importing provider classes directly when you want to swap providers via config. The provider prefix selects the integration package (anthropic, openai, google_genai, bedrock, ...); install the matching langchain-<provider> package alongside. Both forms return a BaseChatModel with the same Runnable interface. See references/models.md and references/setup.md.
LCEL (LangChain Expression Language)
Compose Runnables with |. prompt | model | parser is a Runnable itself — call .invoke({...}), .batch([...]), .stream(...), .ainvoke(...). Use RunnableParallel to fan out, RunnablePassthrough.assign(...) to add fields, RunnableLambda to wrap plain functions. Pass config={"run_name": ..., "tags": [...], "callbacks": [...]} to every top-level call for traceability. See references/lcel.md.
Prompt templates
ChatPromptTemplate.from_messages([("system", "..."), ("human", "{q}")]) produces a Runnable that takes a dict and returns a list of BaseMessage. Use MessagesPlaceholder(variable_name="history") to inject prior turns. Few-shot prompts and prompt-hub pulls are supported. See references/prompts.md.
Structured output
model.with_structured_output(MyPydanticModel) returns a Runnable whose .invoke() yields a parsed Pydantic instance. Under the hood it uses provider-native JSON-schema mode (method="json_schema") or tool calling depending on the provider. Prefer this over manual json.loads of LLM output — schema-enforced decoding catches errors at the model level. See references/structured-output.md.
Tools
Decorate a function with @tool to expose it to a model. The docstring becomes the tool description (the LLM reads it to decide when to call), and type hints become the args schema. For complex args use args_schema=PydanticModel. Bind tools to a model with model.bind_tools([t1, t2]); the model returns AIMessage.tool_calls, you execute them, and feed ToolMessage results back. See references/tools.md.
Retrievers and RAG
Load docs → split with RecursiveCharacterTextSplitter (or MarkdownHeaderTextSplitter for structured text) → embed → store in a VectorStore (Chroma, Qdrant, FAISS, pgvector via langchain-postgres) → vectorstore.as_retriever() → use the Retriever in an LCEL chain. Advanced retrievers: MultiQueryRetriever (LLM rephrases query N times), ContextualCompressionRetriever (re-rank + filter). See references/retrievers-and-rag.md.
Agents
In v1, build agents with from langchain.agents import create_agent. This replaces the deprecated AgentExecutor and the prior langgraph.prebuilt.create_react_agent. create_agent returns a LangGraph-backed runnable that supports tool calling, system prompts, and middleware. For custom graph topologies, drop down to the langgraph skill. See references/agents.md.
Streaming
.stream() / .astream() yields token chunks for the leaf model call. .astream_events() yields semantic events for the full LCEL chain — on_chat_model_stream, on_retriever_end, on_tool_start, etc. — with metadata. Use callbacks (BaseCallbackHandler subclass) for production loggers. See references/streaming.md.
Memory and history
RunnableWithMessageHistory wraps a chain with a BaseChatMessageHistory (in-memory, RedisChatMessageHistory, SQLChatMessageHistory). For new code building agents, prefer a LangGraph checkpointer (InMemorySaver, RedisSaver, PostgresSaver) — it persists the full graph state, not only messages, and integrates with create_agent. Summarization memory classes from v0 are deprecated. See references/memory.md.
Observability with LangSmith
Set LANGSMITH_TRACING=true, LANGSMITH_API_KEY=..., optionally LANGSMITH_PROJECT=.... Every Runnable run auto-traces. Use config={"run_name": "...", "tags": [...]} to label runs. Token usage and latency flow into LangSmith without extra code. See references/observability.md.
Caching
set_llm_cache(InMemoryCache()) for unit tests; SQLiteCache(database_path="...") for local dev; RedisSemanticCache(embedding=..., redis_url=...) for production semantic dedup. Caching only helps for deterministic prompts (low temperature, no time-varying context). See references/caching.md.
Migration from v0
Chain subclassing → LCEL composition. AgentExecutor(...) → create_agent(model, tools, system_prompt=...). chain(input) (calling __call__) → chain.invoke(input). ChatOpenAI(model_name=...) still works but prefer init_chat_model("openai:...") for provider-agnostic code. ConversationBufferMemory → LangGraph checkpointer. See references/migration-from-v0.md.
Behavioral Traits
Always initialize models via init_chat_model("provider:model-id") unless the project explicitly needs provider-specific constructor args
Always use with_structured_output(PydanticModel) for typed output — never json.loads(response.content) over a free-form prompt
Always pass config={"run_name": "...", "tags": [...]} on top-level .invoke / .stream calls so LangSmith traces are meaningful
Always use .invoke() / .ainvoke() — chain(input) (i.e., __call__) is the v0 API path and is deprecated
Prefer .batch([...]) over a Python loop of .invoke — batch uses provider concurrency and shares config
Prefer .ainvoke / .astream in async paths; never mix sync .invoke inside an async event loop without asyncio.to_thread
Set temperature=0 for deterministic / cacheable / test paths; raise only for creative tasks
Use @tool with a clear docstring — the LLM picks tools by docstring, not by function name
Use RecursiveCharacterTextSplitter as the default text splitter; switch to MarkdownHeaderTextSplitter for markdown
When building agents, reach for create_agent first; drop to raw LangGraph only when you need custom state or branches
Stream via astream_events when the UI needs to show intermediate steps (retrieval done, tool called); use stream only when raw tokens are enough
Important Constraints
NEVER subclass Chain for new code — the class hierarchy is retired in v1; compose Runnables with | instead
NEVER use AgentExecutor for new code — it is deprecated; use langchain.agents.create_agent (which sits on LangGraph)
NEVER parse LLM JSON output by hand when with_structured_output is available — schema-enforced decoding is provider-native and far more reliable
NEVER call chain(input) (Runnable __call__) in new code — use .invoke(input); __call__ is the v0 surface and is being removed
NEVER pass plain dicts where the chain expects messages — ChatPromptTemplate produces BaseMessage objects; bypassing the template breaks MessagesPlaceholder and history
NEVER call sync .invoke from inside an async handler without asyncio.to_thread — it blocks the event loop
NEVER assume the embedding model matches the vector store dimension — a mismatch raises at query time, not at insert time; check Embeddings.embed_query("x") length up front
NEVER trace production traffic to LangSmith without setting LANGSMITH_PROJECT per environment — dev runs pollute the prod project
ALWAYS install matching partner packages (langchain-anthropic, langchain-openai, langchain-postgres, ...) — they live outside the langchain core wheel
ALWAYS pin langchain and langchain-core together — minor version skew between them causes ImportError on internal APIs
Related Skills
Parent
python — language baseline (3.11+ comfort, type hints, async)
Sibling — typed schemas
pydantic — with_structured_output is Pydantic v2 under the hood; tool args schemas are Pydantic
Sibling — model SDKs
claude-api — Anthropic Python SDK directly, when you don't need orchestration
