AI / LangGraph LangChain Interview questions
What are common chain composition patterns?
Beyond simple prompt | model | parser pipes, a handful of patterns appear repeatedly in production LangChain applications:
- RAG pattern — retrieve relevant documents, inject them into a prompt, generate an answer. The retriever and passthrough run in parallel so both context and question reach the prompt:
{context: retriever, question: RunnablePassthrough()} | rag_prompt | llm | parser - Router / conditional branch — use
RunnableBranchor a lambda to route different inputs to different sub-chains. Useful for multi-intent chatbots where a general question goes to one chain and a SQL query goes to another. - Map-reduce — split a long document into chunks, process each chunk in parallel with
.batch(), then reduce the results with a combine chain. Standard pattern for summarising books or analysing large codebases. - Refine — process chunks sequentially, passing the previous summary into the next iteration to progressively build a better answer. More accurate than map-reduce for certain summarisation tasks.
- Fallback chain — primary chain with a backup:
gpt4_chain.with_fallbacks([gpt35_chain]). If the primary raises an exception, the fallback is tried automatically. - Branching + merge — run parallel branches (e.g. extract entities, summarise, classify sentiment) and merge their outputs into a final dict for downstream use.
In the standard RAG chain pattern, why does RunnableParallel wrap the retriever and RunnablePassthrough together?
When would you choose a Refine pattern over a Map-Reduce pattern for document summarisation?
More Related questions...
