Add classifier_routing example: route nodes without an LLM call

[jamsea]

Summary

Adds a new example, examples/classifier_routing.py, showing how to route between Flow nodes using a classifier instead of an LLM call.

Some node decisions are trivial: a verification step that only needs a yes/no, a menu where the user picks one of a few known options. Paying for an LLM inference to make that decision adds latency and cost you don't need. This example demonstrates the idiomatic Flows pattern for making those decisions with a plain classifier (a regex here, but the classify() function is the seam to swap in a rules engine or a small binary classifier), while reserving the LLM only for inputs the classifier can't confidently handle.

How it works

The example is a tiny age-verification bot:

• The verify node uses respond_immediately=False, so entering it triggers no LLM completion. It speaks its canned question via a tts_say pre-action and then waits for the user.
• In a normal pipeline, the user's turn-end emits an LLMContextFrame that the LLM service consumes and responds to. That is the LLM call we want to skip.
• ClassifierRouter is a custom FrameProcessor placed between the user context aggregator and the LLM. While in the verify node it classifies the user's answer and either:
  • routes deterministically (yes -> confirmed, no -> declined) via set_node_from_config(...) and drops the LLMContextFrame so the LLM never runs, or
  • passes the frame through to the LLM when it can't classify the input, letting the LLM improvise an answer (the follow-up / off-topic case).
• The terminal confirmed/declined nodes also avoid the LLM: they speak a canned line and end the call with an end_conversation pre-action.

ClassifierRouter drops exactly one frame type (LLMContextFrame), and only in the deterministic branch, because that frame is precisely the LLM-run trigger. Every other frame is always pushed through.

Changes

• New: examples/classifier_routing.py
• examples/README.md: listed the new example under Implementation Patterns
• Changelog fragment

Testing

• ruff format + ruff check: clean
• Offline smoke test: module imports, classify() returns the expected yes/no/None across cases, all three node builders produce valid NodeConfigs, and the router instantiates
• Run it: uv run examples/classifier_routing.py (needs CARTESIA_API_KEY + GOOGLE_API_KEY), then open http://localhost:7860/client. Saying "yes"/"no" routes with no LLM call between turn-end and the reply; asking a question instead falls through to the LLM.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Adds a new example demonstrating an idiomatic “classifier router” pattern to deterministically route between FlowManager nodes without triggering an LLM call, while still allowing an LLM fallback for unclassified inputs.

Changes:

• New examples/classifier_routing.py example implementing a FrameProcessor that drops LLMContextFrame on deterministic routes.
• Updated examples/README.md to list the new example under Implementation Patterns.
• Added a Towncrier changelog fragment documenting the new example.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.

File	Description
examples/README.md	Adds the new example to the examples index.
examples/classifier_routing.py	New runnable example showing classifier-based node routing that can skip LLM inference.
changelog/275.added.md	Changelog entry for the new example.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[jamsea]

Here's a demo:
https://www.loom.com/share/6fcfcfd0aed448948f03ef3b37ebe3b8