llmtoolset 0.4

Tools using ollama and torch that are easy and nice to have for working with local LLMs

llmtoolset

llmtoolset is a Python library designed to provide tools for working with large language models (LLMs) and embeddings without relying on APIs like OpenAI or Hugging Face tokens. It focuses on flexibility and performance using Ollama and PyTorch for embeddings. This toolkit streamlines common operations such as sentence encoding, similarity computation, clustering, and utility functions for LLM-driven workflows.

Installation Requirements

Before installing llmtoolset, ensure you have the correct version of PyTorch installed. Use the command below for CUDA-enabled systems to maximize performance. Optionally, include torchvision and torchaudio for broader PyTorch functionality:

pip install torch --index-url https://download.pytorch.org/whl/cu118

If PyTorch isn't pre-installed, llmtoolset will automatically install a default version of PyTorch.

Features

• Sentence Encoding: Encode text into vector representations using SentenceTransformer.
• Cosine Similarity: Calculate similarity between vectors or batches for clustering, ranking, and comparison.
• Embedding Storage: Save and load embeddings seamlessly in .npy format.
• Nearest Neighbors: Find similar embeddings with customizable thresholds and top-k results.
• Clustering: Group embeddings based on similarity thresholds.
• Tag Extraction: Utility functions to extract and process tags or lists from text.
• Stream Management: Real-time interaction support for LLM streams.

Examples

Encoding Sentences

from llmtoolset.embeddings import SentenceEncoder

encoder = SentenceEncoder()
embeddings = encoder.encode(["This is a test sentence.", "Another sentence to encode."])
print(embeddings)  # Outputs a NumPy array of encoded vectors

Calculating Cosine Similarity

from llmtoolset.embeddings import SentenceEncoder, cosine_similarity

# Initialize the encoder
encoder = SentenceEncoder()

# Define the animals
animals = ["cat", "tiger", "fish"]

# Encode the animal names
embeddings = encoder.encode(animals)

# Calculate cosine similarity between the animals
similarity_cat_tiger = cosine_similarity(embeddings[0], embeddings[1])
similarity_cat_fish = cosine_similarity(embeddings[0], embeddings[2])
similarity_tiger_fish = cosine_similarity(embeddings[1], embeddings[2])

print(f"Cosine Similarity between 'cat' and 'tiger': {similarity_cat_tiger}")
print(f"Cosine Similarity between 'cat' and 'fish': {similarity_cat_fish}")
print(f"Cosine Similarity between 'tiger' and 'fish': {similarity_tiger_fish}")

Generating Tags from Text

from llmtoolset import make_tags

text = "This text discusses machine learning and artificial intelligence."
tags = make_tags(text)
print(tags)  # Example output: ['machine learning', 'artificial intelligence']

Clustering Similar Embeddings

from llmtoolset.embeddings import SentenceEncoder, group_similar_embeddings

# Initialize the encoder
encoder = SentenceEncoder()

# Define the animals
animals = ["cat", "tiger", "lion", "dog", "wolf", "fish", "shark", "whale"]

# Encode the animal names
embeddings = encoder.encode(animals)

# Group similar embeddings
clusters = group_similar_embeddings(embeddings, similarity_threshold=0.5)

# Visually print the clustering
for cluster in clusters:
    print("[Cluster]:")
    for item in cluster:
        print(f"   {animals[item[0]]}")

# Results of this code in testing (tweaking would be needed for perfection)
# [Cluster]:
#    cat
#    tiger
#    lion
#    dog
# [Cluster]:
#    wolf
# [Cluster]:
#    fish
#    shark
#    whale

Stream Interaction

from llmtoolset import activate_stream_printing, deactivate_stream_printing

# Enable real-time stream printing
activate_stream_printing()

# Disable it when no longer needed
deactivate_stream_printing()