langchain

Sub-Skill: LangChain / Agent Framework Conventions

Purpose: Prevent common LangChain mistakes — deprecated chain classes, missing retry/timeout guards, unsafe prompt handling, and unobservable pipelines.

---

Rules

Chain Construction

1. Use LCEL pipe syntax. Always use the LCEL pipe operator (|) to compose runnables instead of deprecated constructor-based chain classes (LLMChain, SequentialChain, TransformChain). Reference: ERR-2026-026
2. Avoid legacy chain imports. Never import from langchain.chains.llm or langchain.chains.sequential; use langchain_core.runnables and langchain_core.prompts instead.
3. Prefer RunnablePassthrough for identity steps. Use RunnablePassthrough to thread context through a chain without mutation rather than writing a lambda that returns its input unchanged.
4. Use RunnableParallel for fan-out. Prefer RunnableParallel over manually calling multiple chains and merging dicts; it expresses intent and enables parallel execution.

Prompts & Output Parsers

1. Use typed output parsers. Always attach an output parser (PydanticOutputParser, JsonOutputParser, StrOutputParser) to chains that produce structured data; never parse raw LLM strings manually downstream.
2. Inject format instructions via partial. Use prompt.partial(format_instructions=parser.get_format_instructions()) to bind parser instructions into the prompt template rather than hard-coding them in the template string.
3. Separate system and human messages. Use ChatPromptTemplate.from_messages([("system", ...), ("human", ...)]) instead of a single PromptTemplate for chat models; mixing roles in one string breaks structured output.

Chat Models vs LLMs

1. Prefer ChatModel over LLM. Always use ChatOpenAI, ChatAnthropic, or equivalent chat-model classes for new code; the base OpenAI LLM class is deprecated for most use cases and lacks tool-calling support.
2. Pin model name explicitly. Never rely on the default model name in a chat model constructor; always pass model="gpt-4o" (or equivalent) so upgrades are intentional.