Workflow Patterns in LLM Systems

Static, reusable structures that chain, route, or parallelize LLM steps. Implemented with LangGraph's StateGraph, nodes, and edges (invoke() / ainvoke()).

Pattern	Essence
1. Sequential Chaining	Step A → Step B → Step C
2. Routing	One input → one of many paths
3. Parallelization	Fan-out → process → aggregate

---

1. Sequential Chaining

Step A → Step B → Step C. Steps run in order; each node's output is the next node's input. No branches—linear pipeline: Node 1 → Node 2 → … → END. State: e.g. original_text, summary, translation, current_step.

Use cases: Document writing (outline → verify alignment → generate); multi-step data processing (extract → transform → summarize).

Implementation:Sequential Chaining/langgraph_chaining.py

---

2. Routing

One input → one path. A router (rule-based or LLM) inspects input and chooses one branch—e.g. Billing / Operations / Technical Support for tickets, or small vs. large model by query complexity. State: query, decider/route, result fields. Use add_conditional_edges to map decision to next node.

Use cases: Customer support routing; model selection (simple → small model, complex → reasoning model).

Implementation:Workflow Routing/langgraph_routing.py

---

3. Parallelization

Fan-out → process → aggregate. Distributor fans out to multiple nodes; each does independent work (e.g. subqueries, document chunks); synthesizer aggregates (e.g. reducer Annotated[list, add]) and optionally calls an LLM for a final summary.

Use cases: RAG with subqueries (parallel retrieval → merge and rank); long-document summarization (summarize chunks in parallel → aggregate).

Implementation:Parallelization/langgraph_parallelization.py

---

Summary

Static workflows fit tasks where the execution path is known in advance. Tools like n8n, Zapier, and UnifyApps offer visual/low-code design for these patterns.

Code:Workflow-Patterns on GitHub