Skip to content

transformers.md

Latest commit

 

History

History
106 lines (73 loc) · 4.57 KB

transformers.md

File metadata and controls

106 lines (73 loc) · 4.57 KB
Raw

Transformer Plugins

Transformers are loaded by IntentService and run on every utterance before pipeline matching begins. There are three transformer stages, each backed by a separate plugin type.

Stages

1. UtteranceTransformersService

Entry point group: opm.utterance_transformer Config key: utterance_transformers

Receives the raw utterance list and may rewrite it. Changes are logged as utterances transformed: X -> Y. Use cases: spelling correction, canonicalisation, language normalisation.

utterances, context = utterance_transformers.transform(utterances, context)

2. MetadataTransformersService

Entry point group: opm.metadata_transformer Config key: metadata_transformers

Receives only message.context and may enrich it with additional metadata. Does not alter the utterance text. Use cases: speaker identification, emotion detection, tagging detected language.

context = metadata_transformers.transform(context)

3. IntentTransformersService

Entry point group: opm.intent_transformer Config key: intent_transformers

Runs after a pipeline match is found. Receives and may modify the IntentHandlerMatch object before the reply is emitted. Use cases: entity normalisation, confidence adjustment, adding context to the match.

match = intent_transformers.transform(match)

Plugin Priority

Chains run in ascending priority order per OVOS-TRANSFORM §4: a plugin with priority = 1 runs first (default 50); later plugins see and may override earlier plugins' output. An explicit "order" list in the config section wins over priorities; loaded plugins absent from the list do not run.

The runner services themselves are the canonical implementations from ovos_plugin_manager.transformer_services (re-exported by ovos_core.transformers); they also implement the OVOS-TRANSFORM §8.1 cancellation contract: a plugin returning "canceled": true + "cancel_reason" stops the chain, and handle_utterance terminates the lifecycle with ovos.utterance.cancelled → ovos.utterance.handled. Full contract → ovos-plugin-manager/docs/transformers.md.

Enabling / Disabling Plugins

Each plugin is enabled or disabled in mycroft.conf under its service config key:

{
  "utterance_transformers": {
    "ovos-utterance-normalizer": {"active": true},
    "my-custom-transformer": {"active": false}
  }
}

A plugin not listed in config is not loaded even if installed.

Split deployments: shared servers can run some of these chains too. ovos-stt-server runs utterance transformers on transcripts, and hivemind-core runs utterance and metadata transformers for text clients. Enable each plugin in exactly one place per deployment, or its effect is applied twice.

Cross-References

Entry point groups

All three transformer types are discovered via ovos-plugin-manager:

Stage Entry point group OPM factory function
Utterance opm.utterance_transformer find_utterance_transformer_plugins()
Metadata opm.metadata_transformer find_metadata_transformer_plugins()
Intent opm.intent_transformer find_intent_transformer_plugins()

→ ovos-plugin-manager/docs/plugin-types.md

Writing transformer plugins

Listener-level transformers (distinct from these)

ovos-dinkum-listener has its own STT-level transformer stage that runs before audio is converted to text. These run post-STT but before recognizer_loop:utterance is emitted: distinct from the three transformer stages here. See ovos-dinkum-listener/docs/transformers.md.

Audio-level transformers

ovos-audio has TTS and dialog transformer stages that run when TTS is synthesised. See ovos-audio/docs/transformers.md.

IntentHandlerMatch

  • IntentHandlerMatch: ovos_plugin_manager.templates.pipeline.IntentHandlerMatch. Fields: match_type, match_data, skill_id, utterance, updated_session. Used by IntentTransformersService. See ovos-plugin-manager/docs/plugin-types.md.

← Pipeline · Home · Next →