An intent-matching pipeline plugin for OpenVoiceOS, built on Model2Vec static embeddings. It classifies an utterance against the intents that loaded skills registered at runtime, either through a pre-trained classifier head or through the skills' own example utterances with no training step. Use it as a semantic fallback behind exact matchers such as Adapt and padacioso.
uv pip install --prerelease=allow ovos-m2v-pipeline{
"intents": {
"ovos-m2v-pipeline": {
"model": "OpenVoiceOS/ovos-m2v-intents-multilingual"
},
"pipeline": [
"ovos-padacioso-pipeline-plugin-high",
"ovos-m2v-pipeline-high",
"ovos-m2v-prototype-pipeline-medium"
]
}
}intents.pipeline replaces the whole pipeline list, not only its m2v
entries; see the pipeline order
for where the other plugins a deployment installs belong in it.
ovos-padacioso-pipeline-plugin-high runs first: a frozen classifier only
answers with a label it was trained on, so an exact template line of a
skill it never saw is claimed earlier, and correctly, by an exact matcher.
See the pipeline order
for the full reasoning.
| Entry-point suffix | Method | Answers from |
|---|---|---|
-high |
match_high() |
the trained head, or the skill's own examples for the standalone prototype plugin |
-medium |
match_medium() |
same, at a lower threshold |
-low |
match_low() |
the skill's own examples by default (low_tier: "prototype") |
| Page | Covers |
|---|---|
| Configuration | Every config key, its default and its type |
| OVOS Pipeline Plugin | Entry points, tiers, modes, caching |
| Models | Published model ids and picking one |
| Text embeddings | The text embeddings plugin and its config |
| Training | Building the corpus and fitting a model |
| Label scheme | Label format, families, dedup and renames |
| Pre-release quirks | Behaviour changes by pre-release version, reset at each stable release |
Training your own model starts at Training.
Apache 2.0. See LICENSE.
First built by TigreGótico for OpenVoiceOS under the ILENIA project, and extended through the NGI0 Commons Fund.
