ozonetel-ai 0.0.13

The Ozonetel AI Library is a powerful tool developed by Ozonetel Communications Pvt Ltd, designed to empower developers with state-of-the-art Artificial Intelligence capabilities. This library provides seamless integration of advanced AI models into your projects, allowing you to leverage the latest in AI technology to enhance your applications.

Ozonetel AI

Overview

The Ozonetel AI project is designed to provide a user-friendly interface for software development using Ozonetel's in-house AI libraries, models, and software solutions. It offers seamless integration with Ozonetel's advanced AI capabilities, allowing developers to harness the power of AI to enhance their applications.

Features

• Text Embedding (Binary Embeddings): The Ozonetel AI project currently offers text embedding functionality, allowing users to convert text into high-dimensional bit vectors for various natural language processing tasks.

Getting Started

To get started with the Ozonetel AI project, follow the steps below:

1. Installation

   pip install ozonetel-ai
   

2. Set Credentials: Before using the text embedding feature, set your credentials by importing the os module and setting the OZAI_API_CREDENTIALS environment variable to point to your credentials file.

   Example:

   import os
   os.environ["OZAI_API_CREDENTIALS"] = "./cred.json"
   

   or,

   credential = {"username":"", "bearer_token":""}
   

3. Text Embedding Extraction Text embedding converts textual data into numerical representations, aiding natural language processing tasks. By capturing semantic meaning, it enhances sentiment analysis, document classification, and named entity recognition. Efficient and transferable, embeddings facilitate faster computation and enable machine learning models to better understand and process text. BinarizeSentenceEmbedding binarizes base embeddings and represents in bits.

   Example:

   # Import `BinarizeSentenceEmbedding` class from the `ozoneai.embeddings` module.
   from ozoneai.embeddings import BinarizeSentenceEmbedding
   
   # Extract Embeddings: Use the `binarize` method to obtain binarized embeddings for given texts .
   # Supported models encoders are `sentence-transformers/paraphrase-multilingual-mpnet-base-v2` and `BAAI/bge-m3`
   # Alternatively if you have stored these models in local directory you can use like `/path/to/paraphrase-multilingual-mpnet-base-v2` or `/path/to/bge-m3`
   
   # Note: for safest use `OZAI_API_CREDENTIALS`. In that case `credential` argument can be ignored. if `credential` is not defined `OZAI_API_CREDENTIALS` will be considered.
   binary_encoder = BinarizeSentenceEmbedding(endcoder_modelid="BAAI/bge-m3", credential = credential)
   
   emb = binary_encoder.encode(["Try me Out"])
   emb_binarized = binary_encoder.get_binary_embeddings(emb, model="sieve-bge-m3-en-aug-v1") # max limit 20 vectors per request
   
   # Access Embedding Attributes: Retrieve various attributes of the embedding object, such as bits, unsigned binary, and signed binary.
   # Get binary representation
   embeddings = emb_binarized.embedding
   

   If you want to check available embeddings you can do it as follows

   from ozoneai.embeddings import list_models
   
   list_models()
   

Examples

Search And Index:

import numpy as np, os
from ozoneai.embeddings import BinarizeEmbedding

from sentence_transformers import SentenceTransformer

# define credentials
# os.environ["OZAI_API_CREDENTIALS"] = "./cred.json"
credential = {"username":"", "bearer_token":""}

base_model = SentenceTransformer("BAAI/bge-m3")
binary_encoder = BinarizeEmbedding(endcoder_modelid="BAAI/bge-m3",credential=credential) 

# Documents to be Indexed
docs = [
    "The cat sits outside",
    "A man is playing guitar",
    "I love pasta",
    "The new movie is awesome",
    "The cat plays in the garden",
    "A woman watches TV",
    "The new movie is so great",
    "Do you like pizza?",
    "Artifical intelligence is all about enabling computers to simulate human capabilities.",
    "Artifical intelligence is there to help human",
    "Language moodel has lot of impact on artifical intelligence",
    "Artifical intelligence is reigning in industry",
    "ROI is one the useful measure for any investors",
    "Understand before you invest"
]

# Extract Embeddings for documents to be indexed
docs_emb = base_model.encode(docs)

# Binarise embeedings
docs_emb_binarized = binary_encoder.get_binary_embeddings(docs_emb, model="sieve-bge-m3-en-aug-v1")

# Searching

# query_text = "the girl was sitting in the park"
query_text = "I love Artifical Intellegence"
# query_text = "Artifical intellegence helps optimising software"
# query_text = "what is machine learning"

# Extract Embeddings query

# Compute sentence embedding using SentenceTransformers (i.e. bge-m3 or paraphrase-multilingual-mpnet-base-v2)
query_emb = base_model.encode([query_text])

# Binarise embeedings
query_emb_binarized = binary_encoder.get_binary_embeddings(query_emb, model="sieve-bge-m3-en-aug-v1")

# compute distance of docs from query
topn = 5
dist_pbit = np.bitwise_xor(docs_emb_binarized.embedding, query_emb_binarized.embedding)
dist_bit = np.unpackbits(dist_pbit, axis=1)
dist = np.sum(dist_bit, axis=1)
topn_indices = np.argsort(dist)[:topn]
scores = (1 - (dist/(query_emb_binarized.embedding.shape[1]*8))).round(3)

# Print search result in order
for i, doci in enumerate(topn_indices):
    print(f"{i}. {docs[doci]} [{doci}] ({scores[doci]})")

try more..

Benchmarks

Classification

S.N	Dataset	paraphrase-multilingual-mpnet-base-v2	siv-sentence-bitnet-pmbv2-wikid-small	bge-m3	sieve-bge-m3-en-aug-v0	sieve-bge-m3-en-aug-v1
1	Amazon Counterfactual Classification(en)	75.81	79.06	75.63	79.23	78.28
2	Amazon Polarity Classification	76.41	70.19	91.01	86.81	85.69
3	Amazon Reviews Classification	38.51	34.29	46.99	43.51	43
4	Banking 77 Classification	81.07	75.89	81.93	82.06	82.75
5	Emotion Classification	45.83	40.26	50.16	42.34	42.4
6	Imdb Classification	64.57	61.14	87.84	85.06	84.44
7	Massive Intent Classification (en)	-	65.6	71.08	68.9	70.22
8	Massive Scenario Classification (en)	-	70.37	76.64	71.29	72.68
9	MTOP Domain Classification (en)	89.24	87.22	93.36	87.56	88.37
10	MTOP Intent Classification (en)	68.69	69.45	66.58	74.11	74.02
11	Toxic Conversations Classification	71.02	70.26	72.6	68	68.59
12	Tweet Sentiment Extraction Classification	59.03	54.49	63.71	56.96	56.91

STS

S.N	Dataset	paraphrase-multilingual-mpnet-base-v2	siv-sentence-bitnet-pmbv2-wikid-small	bge-m3	sieve-bge-m3-en-aug-v0	sieve-bge-m3-en-aug-v1
1	BIOSSES	76.27	65.29	83.38	82.91	83.97
2	SICK-R	79.62	76.01	79.91	75.26	76.5
3	STS12	77.9	71.25	78.73	66.95	68.88
4	STS13	85.11	78.4	79.6	64	69.09
5	STS14	80.81	74.23	79	62.83	67.74
6	STS15	87.48	81.41	87.81	80	82.08
7	STS16	83.2	79.13	85.4	77.87	79.31
8	STS17(en-en)	86.99	85.4	87.13	84.1	85.32
9	STS Benchmark	86.82	81.34	84.85	74.13	77.19

License

The Ozonetel AI project is licensed under the MIT License. Please refer to the LICENSE file for more information.