Workflow Patterns

Building reliable AI systems often involves standard architectural patterns.

1. Linear Chains

The simplest workflow. Step A -> Step B -> Step C. Example: Translate user input -> Summarize translation -> Extract keywords.

2. Routing

Directing an input to the specialized workflow best suited to handle it. Example: If the user asks about "Refunds", route to the Support Workflow. If they ask about "Coding", route to the Technical Workflow.

3. Parallelization

Running multiple tasks at once to save time or generate diverse perspectives.

• Sectioning: Breaking a task (like writing a report) into sections and generating them in parallel.
• Voting: Generating 3 distinct answers and having a model pick the best one.

4. Evaluator-Optimizer

A feedback loop where one model generates content and another (or the same one) evaluates and critiques it.

1. Generate: Draft an email.
2. Evaluate: Check if it's polite and clear.
3. Refine: Rewrite based on the feedback.
4. Repeat until it passes.

5. Map-Reduce

Useful for processing large documents.

1. Map: Summarize every chunk of a long document independently.
2. Reduce: Combine all the summaries into a final overview.