A Mental Model: Everything Is a Pipeline of Typed Steps

Input, transform, output

Both frameworks reward the same way of thinking: an LLM app is a pipeline where each step takes a value and returns a value. A prompt template turns a dict into messages; a model turns messages into a message; a parser turns a message into a Python object; a retriever turns a query into documents; a tool turns arguments into a result. When you know each step's input and output type, you can swap parts, test them separately (with fakes) and see exactly where a bug appears. Keep side effects (sending email, writing to a database) in clearly marked steps that can be approved or disabled.

The pipeline view

Types in brackets show what flows between steps.

{question: str} -> [prompt] -> messages -> [model] -> AIMessage -> [parser] -> str / object

retrieval:  str -> [retriever] -> list[Document] -> [format] -> str -> (into the prompt)
tools:      AIMessage(tool_call) -> [tool] -> ToolMessage -> back to the model

Quick check: Why think about each step's input and output types?

• It makes the model smaller
• It lets you swap, test and debug steps independently
• It removes the need for prompts
• It speeds up the GPU