Pipeline Recipes
Fully worked end-to-end pipelines for common domains, with prompts, schemas, and decomposition rationale.
Each recipe below is a complete flow you can adapt: tree shape, block prompts, output schemas, sample API call, and — most importantly — the reasoning behind why the pipeline is shaped the way it is. If you're an LLM building a flow for a user, these are reference templates.
Read Flows & Blocks and Pipeline Patterns first.
Recipes
- Essay Feedback for Language Learners — detect → grade → translate feedback into the learner's native language.
- Support Ticket Triage — classify intent, extract entities + sentiment in parallel, draft a grounded response.
- Document Q&A with Citations — extract relevant passages, answer only from them, cite by paragraph.
- Content Moderation — parallel independent checks, deterministic verdict.
Picking a Recipe
| If your pipeline is roughly… | Start from |
|---|---|
| One LLM call wrapped in business logic | Single Block |
| Detect → process → translate / format | Essay Feedback |
| Classify → enrich → generate response | Support Ticket Triage |
| Search → answer with citations | Document Q&A with Citations |
| Independent checks → deterministic verdict | Content Moderation |
When your use case doesn't match cleanly, reach back to Pipeline Patterns for the structural primitives and compose from there.