LangChain

• langchain.*.invoke(...)
• ChatOpenAI(...), ChatAnthropic(...), ChatVertexAI(...), ChatBedrock(...), etc.
• LLMChain(...)

{{PROVIDER_ASSERTS}} substitution

LangChain is a meta-framework — it wraps a specific provider. Walk one level deeper: the chat-client class names the provider. Read the user’s function (and immediate imports) to identify which Chat* class is instantiated, then emit the assert for THAT provider per the table below:

Emit the assert for the underlying provider, not LangChain itself. Example: if the user’s function has from langchain_anthropic import ChatAnthropic, emit:

If the call-site uses multiple chat clients (rare), emit asserts for each.

Adapter notes

• chain.invoke({...}) returns either a string (for simple chains) or an AIMessage (for chat-based chains). Extract .content if it’s the latter.
• LangChain LCEL chains (prompt | llm | parser) return the parser’s output type directly — usually a string. Trust the user’s function signature.
• Async: chain.ainvoke(...) — wrap with asyncio.run(...).

Common gotchas

• LangChain configures provider via env vars by default but ALSO supports per-instance kwargs (ChatOpenAI(api_key="sk-...")). If the user’s function passes a key explicitly, the env var is irrelevant — emit a # Note: comment instead of an assert.
• LangChain prompts/ directory conventions vary widely; the wrapped function should encapsulate prompt loading, so task_fn just calls function(input_data) and trusts the chain.