Caching Steps in YAML

This page shows how to cache step results in YAML using kind: cache, and when caching is appropriate.

Overview

• kind: cache wraps an inner step and caches its output keyed by input.
• Use caching for expensive, deterministic operations where identical inputs recur (LLMs, remote APIs).

Basic Syntax

- kind: cache
  name: cached_stringify
  wrapped_step:
    kind: step
    name: stringify
    agent: { id: "flujo.builtins.stringify" }

Example: Caching an LLM Call

- kind: cache
  name: cached_summarize
  wrapped_step:
    kind: step
    name: summarize
    uses: agents.summarizer
    input: "{{ context.article_text }}"

Best Practices

• Ensure the wrapped step is deterministic with respect to its input.
• Avoid caching steps that rely on external state unless such state is encoded into the input.
• Start with the default cache backend; add backend configuration later if needed.

Troubleshooting

• If you don't observe hits on repeated runs with identical inputs, verify the wrapped step input really matches (no hidden fields or timestamps).
• For debugging, inspect step_history[-1].metadata_["cache_hit"] — it should be True for cache hits.