ovos-markov-pipeline-plugin

An OVOS intent pipeline plugin that classifies utterances with an ensemble of per-intent Markov chains, scored by perplexity.

It trains one word-level Markov chain per intent from example utterances. At query time it computes the perplexity of the utterance under every model and picks the intent whose model is least "surprised": the lowest perplexity wins. Confidence comes from perplexity, so the plugin fits into the OVOS high, medium, and low confidence pipeline tiers.

Why a Markov pipeline?

Most OVOS intent matchers either pattern-match (Adapt, padacioso) or train a neural or SVM classifier (padatious, linha-fina). A Markov-chain language model sits in between. It is a lightweight statistical model that learns the word order of each intent's sample set. It trains in milliseconds, needs no GPU, and degrades gracefully with very few samples. It is a practical baseline for small skill sets and a useful comparison point against the other matchers.

Install

pip install ovos-markov-pipeline-plugin

Optional stemming support (recommended for non-English use):

pip install ovos-markov-pipeline-plugin[stem]

Quick example

Enable the pipeline in mycroft.conf:

{
  "intents": {
    "pipeline": [
      "ovos-markov-pipeline-plugin-high",
      "ovos-markov-pipeline-plugin-medium",
      "ovos-markov-pipeline-plugin-low"
    ]
  }
}

Skills register their intents over the bus exactly as they do for padatious. No skill-side changes are required.

Documentation

Full documentation lives in docs/:

• Concepts: how Markov-chain perplexity matching works
• Quickstart: install, configure, run
• Configuration: every config key explained
• Pipeline integration: OPM entry point and bus protocol
• Entity extraction: HMM BIO slot tagging
• Calibration: measuring and tuning confidence
• Model caching: ONNX disk cache
• Tuning: order, smoothing, stemming, fallbacks
• Troubleshooting: common issues and FAQ

License

Apache-2.0