| name | ml-component-integration |
| description | Integrate a new ML component (classifier, reranker, embedder, etc.) into an existing pipeline by locating insertion points, implementing the module, and wiring it in. |
/ml-component-integration
When to use
Use this when adding a new ML model or inference component (e.g., intent classifier, reranker, embedder) to an existing processing pipeline. Applies whenever the component needs a defined insertion point, a standalone module, and a hook into existing logic.
Steps
- Identify insertion points — Read the pipeline code to find exactly where the new component should receive input and emit output (e.g., between two existing processors or inside a search/retrieval function).
- Confirm model contract — Check
train.py or equivalent to pin down the model architecture, input shape, output format, and artifact schema (weights file format, label names, etc.).
- Implement the component module — Write the standalone module (e.g.,
agent/intent.py, memory/reranker.py) with a clean interface matching the pipeline's data types.
- Hook into existing system — Patch the integration point in the host file (e.g.,
memory/facts.py search()) to call the new module, handling fallback/disabled cases gracefully.
- Verify end-to-end — Confirm the frame/data flows correctly through the new component with a smoke test or log trace.
Examples
- Adding an intent classifier between
audio_tags and user_agg in a voice pipeline: read TranscriptionFrame after tags, emit intent label before the LLM aggregator.
- Plugging a BM25 + cross-encoder reranker into
memory/facts.py search() as a post-processing step over raw retrieval results.
