Member-only story
MapReduce is a way to process a list of items and combine the results into a single output.
The idea has two steps:
Map: Process each item in the list separately.** Reduce**: Combine all the individual results into one final result.
This pattern is commonly used in distributed systems, but it is also useful in LLM applications.
In an LLM workflow, the map step often means taking a list generated by one model call and running a separate model call (or API call) for each item. The reduce step then combines all those results with a final model call.
In LangGraph, this pattern is implemented using Send. In this article, I will build a simple book recommendation workflow to show how Send distributes work across multiple parallel tasks and then combines the results into a final output.
The Problem with a Fixed Graph #
A standard LangGraph graph has a fixed set of nodes and edges. This works well when the number of steps is known in advance. It becomes difficult when the number of steps depends on the data received at** runtime**.
For example — suppose a user asks for book recommendations on the history of the Football World Cup. The first LLM call may return 2 books, 4 books, or even 10. The graph doesn’t know this until that call finishes.