Langchain

LangChain Python Ecosystem Reference

Use this skill for Python LangChain ecosystem work. It covers package boundaries, model initialization, LCEL/runnables, agent construction, retrieval wiring, LangGraph stateful workflows, LangSmith tracing/evals, and migration away from legacy imports. Read only the reference files needed for the task.

If the user is working on LangChain.js / LangGraph.js / JS/TS, do not reuse Python imports or package advice from this skill. Go to the JS/TS docs instead.

Quick Start

1. Identify the owning package first: langchain, langchain-core, a provider package, langgraph, langsmith, or legacy langchain-classic.
2. Open the single best reference file from references/.
3. Add a second reference only when the task crosses a real boundary, such as create_agent + LangGraph persistence or RAG + provider packages.
4. Verify imports and pip package names before writing code.
5. State which reference files you used and what tests or smoke checks are required.

When Not To Use This Skill

• LangChain.js / LangGraph.js — This skill is Python-only.
• Generic provider SDK work without LangChain — Use provider-native docs when the code does not use LangChain abstractions.
• Pure vector database administration — Cluster ops, index tuning, or deployment for Qdrant/Chroma/etc. are out of scope unless the task is specifically LangChain integration wiring.
• Framework lifecycle questions — Pair the relevant framework skill when request lifetimes, background jobs, or server startup/shutdown behavior belong to Django, Litestar, FastAPI, or another framework.

Critical Rules

1. Do not treat the ecosystem as one package — langchain, langchain-core, provider integrations, langgraph, langsmith, and langchain-classic have different roles.
2. langchain-core owns abstractions, not integrations — Its own source says: "No third-party integrations are defined here." Do not invent provider imports from core.
3. Provider integrations live in separate packages — Examples verified from source: langchain-openai, langchain-anthropic, langchain-chroma, langchain-qdrant.
4. Prefer main langchain for current agent/app work — The langchain-classic package is explicitly for legacy chains, community re-exports, indexing API, deprecated functionality, and more.
5. Use create_agent for standard tool-loop agents — LangChain recommends LangGraph only when you need heavier customization, deterministic + agentic orchestration, or carefully controlled state/latency.
6. Use LangGraph when durability or human intervention is required — StateGraph is the low-level orchestration framework for long-running, stateful workflows.
7. StateGraph must be compiled before execution — The source explicitly says the builder cannot execute until .compile() is called.
8. InMemorySaver is not a production persistence strategy — LangGraph documents it for debugging or testing and recommends a durable saver such as Postgres for production.
9. LangSmith is observability/evals, not runtime orchestration — Use it for tracing, datasets, benchmarking, pytest-based evaluation, and monitoring.
10. Do not guess provider params or model IDs — init_chat_model() requires the provider package to be installed and recommends exact model IDs from provider docs.
11. Treat configurable_fields="any" as a security risk — The source explicitly warns that runtime config can alter api_key, base_url, and other sensitive fields.
12. Legacy snippets need migration review before reuse — Old examples may point at langchain-classic, deprecated LangGraph types, or outdated imports.

Reference Map

Task Routing

• Import error, install question, package confusion -> references/package-map.md
• Model initialization, provider string, configurable model, LCEL pipeline -> references/models-prompts-runnables.md
• Standard tool-calling agent, middleware, structured output -> references/agents-tools-structured-output.md
• RAG / retrieval / chunking / vector store wiring -> references/retrieval-integrations.md
• Long-running workflow, persistence, interrupts, subgraphs, prebuilt nodes -> references/langgraph-stateful-workflows.md
• Tracing, datasets, benchmark evals, pytest integration -> references/langsmith-observability-evals.md
• Old tutorial, langchain-classic, deprecated imports, migration review -> references/migration-classic-gotchas.md

Output Expectations

• Name the reference files used.
• Name the owning package(s) explicitly.
• Call out any separate pip packages that must be installed.
• State whether the task belongs in plain LangChain, LangGraph, or LangSmith.
• If the answer depends on current package generation, say so plainly instead of pretending old blog posts are current.
• State the minimum verification step: import smoke test, runnable/agent smoke test, LangGraph checkpoint test, or LangSmith trace/eval smoke test.

Content Ownership

This skill owns the Python LangChain ecosystem: langchain, langchain-core, provider integration packages, langchain-text-splitters, langgraph, langsmith, current agent APIs, LCEL/runnables, retrieval wiring, tracing/evals, and legacy migration boundaries.

This skill does not own LangChain.js / LangGraph.js or provider-native SDK design outside LangChain usage.

Source: dekoza/oh-my-slop — distributed by TomeVault.