nux-ai / vectors

Toolkit designed for developers to evaluate, select, and deploy embedding models. It streamlines the lifecycle from model evaluation to data embedding and querying.

Embedding Model Evaluation & Integration Toolkit

Welcome to the Embedding Model Evaluation & Integration Toolkit, an open-source project designed to streamline the end-to-end lifecycle of embedding model evaluation, index creation, and querying embeddings.

Our mission is to provide a robust and simple-to-use interface for developers to leverage the power of embedding models across various applications, from natural language processing to vector search databases.

🚀 Purpose

The toolkit aims to empower developers by simplifying the process of:

• Evaluating different embedding models to find the best fit for your specific dataset and query patterns.
• Creating efficient indices for fast retrieval.
• Querying embeddings to unlock insights and patterns within your data.

📘 Lifecycle Walkthrough

1. Evaluate and Select a Model

Jumpstart your project by evaluating potential models against your data and criteria.

# /utilities/evaluate.py
evaluate_instance = Evaluate(
    model="all-MiniLM-L6-v2", 
    testing_set=[{'text': 'Sample text 1'}, {'text': 'Sample text 2'}]
)

evaluate_instance.evaluate(
    query="Example query", 
    acceptance_criteria={['Sample text 2', 'Sample text 1'],
    order="specific"
)

2. Mount the Selected Model(s) via HTTP

Easily integrate models into your workflow with HTTP endpoints.

Mount the model:

curl -X POST http://localhost:5000/mount_model \
     -H "Content-Type: application/json" \
     -d '{"model_name": "all-MiniLM-L6-v2"}'

Retrieve an embedding:

curl -X POST http://localhost:5000/get_embedding \
     -H "Content-Type: application/json" \
     -d '{"text": "Example text for embedding."}'

3. Create Your Index

Optimize data retrieval with custom indices tailored to your model's embeddings.

# /cloud/mongodb.py
atlas = Atlas(field_names_and_dims, "index_keyword_map_test")
atlas.create_index()

4. Load Data Using Selected Model

Embed and store your data efficiently using the model of your choice.

# /utilities/load.py
data_loader = DataLoader("your_db_name", "your_collection_name")
data_loader.load(data_mapping)

5. Design Query and Evaluate Results

Unleash the full potential of your data with powerful querying capabilities.

client.collection.aggregate([
  {
    '$vectorSearch': {
      'index': 'default',
      'path': 'plot_embedding_384',
      'queryVector': 'lorem ipsum',
      'numCandidates': 150,
      'limit': 10
    }
  },
  {
    '$project':
    {
      'plot': 1, 
      'title' : 1
    }
  }
])

🗺 Library Roadmap

We're constantly looking to expand the toolkit's capabilities, with plans to include:

• Each time an embedding model is changed:
  • Spark job to paralellize re-embedding
  • Migration of previous vectors to S3
• Migration of vectors from other stores (ex Pinecone to Mongo)
• Federated KNN querying capabilities.
• Containerization of embedding models for ease of deployment.

🌟 Why Contribute?

Contributing to this toolkit not only helps improve a project at the forefront of embedding technology but also connects you with a community of like-minded developers. Whether you're looking to:

• Enhance your understanding of embedding models and their applications.
• Share your expertise and learn from others in the field.
• Drive innovation in embedding model evaluation and integration.

We welcome contributions of all forms, from code improvements and feature additions to documentation and examples!

🛠 How to Contribute

1. Fork the repository: Start with a personal copy of the project.
2. Pick an issue or propose a feature: Look for open issues or suggest new ideas.
3. Submit a pull request: Implement your changes and submit a PR for review.