When a filter cloth 🏳️ is needed rather than a simple RAG 🏴‍☠

Deployment, Stats, & License

---

What is it?

Muzlin merges classical ML with advanced generative AI to efficiently filter text in the context of NLP and LLMs. It answers key questions in semantic-based workflows, such as:

• Does a RAG/GraphRAG have the right context to answer a question?
• Is the topk retrieved context too dense/sparse?
• Does the generated response hallucinate or deviate from the provided context?
• Should new extracted text be added to an existing RAG?
• Can we detect inliers and outliers in collections of text embeddings (e.g. context, user question and answers, synthetic generated data, etc...)?

Note: While production-ready, Muzlin is still evolving and subject to significant changes!

Quickstart

1. Install Muzlin using pip:

   pip install muzlin

2. Create text embeddings with a pre-trained model:

   import numpy as np
   from muzlin.encoders import HuggingFaceEncoder # Ensure torch and transformers are installed
   
   encoder = HuggingFaceEncoder()
   vectors = encoder(texts)  # texts is a list of strings
   vectors = np.array(vectors)
   np.save('vectors', vectors)

3. Build an anomaly detection model for filtering:

   from muzlin.anomaly import OutlierDetector
   from pyod.models.pca import PCA
   
   vectors = np.load('vectors.npy')  # Load pre-saved vectors
   
   od = PCA(contamination=0.02)
   
   clf = OutlierDetector(mlflow=False, detector=od) # Saves joblib moddel
   clf.fit(vectors)

4. Filter new text using the trained model:

   from muzlin.anomaly import OutlierDetector
   from muzlin.encoders import HuggingFaceEncoder
   import numpy as np
   
   clf = OutlierDetector(model='outlier_detector.pkl')  # Load the model
   encoder = HuggingFaceEncoder()
   
   vector = encoder(['Who was the first man to walk on the moon?'])
   vector = np.array(vector).reshape(1, -1)
   
   label = clf.predict(vector)

Integrations

Muzlin integrates with a wide array of libraries for anomaly detection, vector encoding, and graph-based setups.

Anomaly Detection	Encoders	Vector Index
Scikit-Learn PyOD (vector) PyGOD (graph) PyThresh (thresholding)	HuggingFace OpenAI Cohere Azure Google Amazon Bedrock Fastembed Mistral VoyageAI	LangChain LlamaIndex

Simple Schematic Implementation

---

Resources

Example Notebooks

Notebook	Description
Introduction	Basic semantic vector-based outlier detection
Optimal Threshold	Selecting optimal thresholds using various methods
Cluster-Based Filtering	Cluster-based filtering for question answering
Graph-Based Filtering	Using graph-based anomaly detection for semantic graphs like GraphRAG

What Else?

Looking for more? Check out other useful libraries like Semantic Router, CRAG, and Scikit-LLM

---

Contributing

Muzlin is still evolving! At the moment their are major changes being done and the structure of Muzlin is still being refined. For now, please leave a bug report and potential new code for any fixes or improvements. You will be added as a co-author if it is implemented.

Once this phase has been completed then ->

Anyone is welcome to contribute to Muzlin:

• Please share your ideas and ask questions by opening an issue.
• To contribute, first check the Issue list for the "help wanted" tag and comment on the one that you are interested in. The issue will then be assigned to you.
• If the bug, feature, or documentation change is novel (not in the Issue list), you can either log a new issue or create a pull request for the new changes.
• To start, fork the dev branch and add your improvement/modification/fix.
• To make sure the code has the same style and standard, please refer to detector.py for example.
• Create a pull request to the dev branch and follow the pull request template PR template
• Please make sure that all code changes are accompanied with proper new/updated test functions. Automatic tests will be triggered. Before the pull request can be merged, make sure that all the tests pass.