An orchestration system for AI/ML classification at scale.
Maybe you need to comb through a million telephone call transcripts, looking for customers saying, "Take me off your list and don't call me again".
Maybe you need to detect Russian land mines in images.
Maybe you need to rank a million content items in a search engine.
Maybe you need to detect bored telephone operators speaking in a monotone voice in audio recordings.
Building classifiers is easy, but wrangling lots of them at scale is hard. You can find a Jupyter notebook in a few minutes that you can use to train a machine-learning model. But you need to get your training data into it. You'll need to fiddle with your model's parameters to get good performance. You'll need to be able to evaluate it to see how well it's doing. You'll need to be able to deploy it to production and operate it. You'll need to make notes somewhere about how you did that since you'll have to repeat the process over and over and over.
Plexus is not technology for helping you build models. It's technology for helping you manage models at scale. In biology, a "plexus" is an interconnected structure that fans out into many branches, and this Plexus is a way to fan out classification on a massive scale through standardization.
To do classification at scale, you'll need a way to select appropriate models from a toolbox, configure setup and hyperparameters for training, handle training and evaluation, and deploy models to production. You'll need a process for implementing a solution for each new classification problem using as many reusable tools as possible. And you'll need to be able to repeat that process over and over again.
Plexus provides some tools built around classic machine-learning models, others builts on agentic processes using sequences of LLM API requests, and others based on numerical or heuristic methods for scoring. And most importantly, it provides a standard framework where you can plug in your own custom models.
Plexus is an opinionated framework for doing that sort of classification work. Among its opinions:
Classifications revolve around "scores", which can be "questions" or yes/no detection, or assigning a scalar value to a piece of content. Or any task that can be reduced to binary or multi-class classification or, generally, "scoring". Scores are grouped into "scorecards".
All mission-critical functionality should always revolve around versioned code and automated testing and deployment pipelines. Including infrastructure.
You can't optimize a metric you're not measuring and tracking. Plexus uses MLFlow during model development for optimizing metrics like accuracy, precision, recall, and F1 score. The overall model lifecycle revolves around the MLFlow model registry.
There are a lot of valuable ideas out there for doing ML classification, and Plexus serves as a curated collection of some of the best models and tools for doing that. For example, the DeepLearningSemanticClassifier is an implementation of Google's BERT For Sequence Classification that supports the HuggingFace Auto Classes, so that you can plug in other embedding models like DistilBERT or RoBERTa. Whereas the FastTextClassifer is a wrapper around Meta's fastText. Plexus provides a growing toolkit of options to help you build scorecards quickly.
Based on the principle of formulating the problem in terms that the model can solve, Plexus orchestrates scoring using large language models using an agentic decision tree and a set of fine-grained prompts for computing the classification in multiple, small steps. It also handles filtering transcripts for relevance and chunking longer text content into smaller pieces and computing an overall score from the chunks.
For the best results, you can also build custom classifiers. Plexus provides a framework for doing that, and it's designed to be easy to use. You can define your own classifier by subclassing the Score class and following the standard Input and Result Pydantic classes. And you can define your own data processors by subclassing the Processor class. You can then use your custom classifier and processor in your scorecard.
To avoid when-you-have-a-hammer-everything-looks-like-a-nail syndrome, Plexus makes it easy to swap out different types of machine-learning or computational or AI classifiers for different scores, so that users can experiment to find the ideal configuration for their own scorecards.
Plexus is designed to run on Apache Spark, a powerful option for processing training or inference or both at scale. [Spark support is still under development]
Plexus is a Python module that you install in your Python project that contains your proprietary code for accessing your data. It uses a scorecards/ folder in your project to define the configuration for your scoring. Each scorecard is represented by a YAML configuration file. For example, if you have a scorecard for QA, you would have a YAML file in scorecards/qa.yaml that defines the scoring for QA, and it might look like this:
name: Call Center QA
key: contactics
metadata:
foreign_id: 012
transcript_column: Transcription
scores:
IVR Answered:
id: 1234
tags:
- sales
- IVR
class: DeepLearningSemanticClassifier
embeddings_model: 'distilbert-base-uncased'
multiple_windows: true
maximum_windows: 8
maximum_tokens_per_window: 512
start_from_end: true
embeddings_model_trainable_layers: 2
batch_size: 20
number_of_epochs: 10
warmup_learning_rate: 0.000000125
number_of_warmup_epochs: 1
plateau_learning_rate: 0.000042
number_of_plateau_epochs: 5
learning_rate_decay: 0.75
early_stop_patience: 3
l2_regularization_strength: 0.001
dropout_rate: 0.05
data:
class: AWSDataLakeCache
queries:
- scorecard-id: 012
score-id: 1234
value: 'Yes'
number: 10000
- scorecard-id: 012
score-id: 1234
value: 'No'
number: 10000
processors:
- class: RemoveSpeakerIdentifiersTranscriptFilter
- class: RemoveStopWordsTranscriptFilter
Do Not Call Request:
id: 5678
tags:
- compliance
class: ExplainableClassifier
top_n_features: 50000
leaderboard_n_features: 10
target_score_name: 'Do Not Call Request'
target_score_value: 'Yes'
ngram_range: "1,4"
scale_pos_weight_index: 1
data:
class: AWSDataLakeCache
queries:
- scorecard-id: 012
score-id: 5678
value: 'Yes'
number: 10000
- scorecard-id: 012
score-id: 5678
value: 'No'
number: 10000
processors:
- class: RemoveSpeakerIdentifiersTranscriptFilter
- class: ExpandContractionsProcessorEach Score implementation must follow the standard, based on the MLFLow standard for model interface. Standard MLFlow models handle inference through a predict() function, and Plexus specifies standard Input and Result Pydantic classes that are extensible for custom models. For example, the standard Score.Input class contains fields for the classification and the confidence of the classification, whereas the ExplainableClassifier extends that to include an explanation field.
Train a whole scorecard:
plexus train --scorecard-name "Call Center QA"
Train an individual score:
plexus train --scorecard-name "Call Center QA" --score-name "IVR Answered"
The training process will produce a trained model and save it to the MLFlow model registry. It will also evaluate the model and log the metrics and artifacts like confusion matrices and ROC curves to MLFlow.
Some classifiers might support evaluation without supporting training, like classifiers built using agentic LLM processes. You can evaluate these classifiers for different metrics, separately from training.
plexus evaluate accuracy --scorecard-name "Call Center QA" --score-name "Subjective Customer Engagement Score"
Plexus is a work in progress, with gaps in test coverage and documentation, and with a rapidly-evolving feature set. Use at your own risk, and please feel free to contact us for help.
Plexus is developed and maintained by Anthus AI Solutions. Please contact us if you need help using Plexus for your classification needs.
Clone this repository. We're not in PyPi yet, so you'll have to install it locally.
pip install -e .
Run all the tests:
pytestRun a single test:
pytest plexus/processors/RelevantWindowsTranscriptFilter_test.py -k test_multiple_keywordsor:
pytest -k test_multiple_keywordsRun tests with coverage:
pytest --cov=plexusGenerate an HTML coverage report:
pytest --cov=plexus --cov-report=htmlThat will generate a folder called htmlcov with an HTML coverage report. One way to view that is to run:
python -m http.serverAnd then go to http://localhost:8000
Run tests with watch:
ptwWe build the Sphinx documentation at https://anthusai.github.io/Plexus/ with the files in the documentation/ folder.
TODO: Instructions for contributing to the project.
Plexus is open-source under the MIT license.