Choosing chain components for the input and output a design calls for

When designing a chain for a Generative AI application, you select components based on what input the model needs and what output format the application requires. This typically involves combining retrievers, prompt templates, LLMs, and output parsers in a pipeline framework like LangChain, orchestrated to transform raw context into a structured or unstructured final response.

1 · Learn the must-know

• 1A prompt template component is required whenever you need to inject retrieved context, chat history, or user variables into a consistent instruction format before sending it to the LLM.
• 2A retriever component (e.g., a vector store retriever) is selected when the desired output requires grounding in external or proprietary documents (RAG pattern), converting a query into relevant context chunks.
• 3An output parser component is needed when the application requires structured output (e.g., JSON, a specific schema, or extracted fields) rather than free-form text from the LLM.
• 4Chain components should match input/output types end-to-end: a retriever's output (documents) must align with what the prompt template expects as input, and the LLM's output must align with what the parser or downstream component expects.
• 5For multi-step or agentic behavior, additional components such as tools or intermediate chains are selected only when the task requires actions beyond a single LLM call, like calling an API or performing calculations.
• 6Choosing unnecessary components (e.g., an output parser when plain text suffices) adds latency and complexity without benefit, so components should be selected to match the minimal desired input/output contract.

2 · Check your understanding