LLMlight 0.5.0

LLMlight is a Python library for ...

LLMlight

LLMlight is a Python package for running Large Language Models (LLMs) locally with minimal dependencies. It provides a simple interface to interact with various LLM models, including support for GGUF models and local API endpoints. ⭐️Star it if you like it⭐️

---

Key Features

Feature	Description
Local LLM Support	Run LLMs locally with minimal dependencies.
Full Prompt Control	Fine-grained control over prompts including Query, Instructions, System, Context, Response Format, Automatic formatting, Temperature, and Top P.
Single Endpoint for All Local Models	One unified endpoint to connect different local models.
Flexible Embedding Methods	Multiple embedding strategies: TF-IDF for structured documents, Bag of Words (BOW), BERT for free text, BGE-Small.
Advanced Retrieval Methods	Supports Naive RAG with fixed chunking and RSE (Relevant Segment Extraction).
Context Strategies	Advanced reasoning for complex queries using Global-reasoning and Chunk-wise approaches.
Local Memory	Video memory storage for efficient local use.
PDF Processing	Native support for reading and processing PDF documents.

---

Documentation & Resources

• Documentation
• Blog Posts
• GitHub Issues

Quick Start

Installation

# Install from PyPI
pip install LLMlight

## Examples


### 1. Check Available Models at Endpoint

```python
from LLMlight import LLMlight

# Initialize client
from LLMlight import LLMlight
# Initialize with LM Studio endpoint
client = LLMlight(model='mistralai/mistral-small-3.2',
                  endpoint="http://localhost:1234/v1/chat/completions")

modelnames = client.get_available_models(validate=False)
print(modelnames)

2. Basic Usage with Endpoint

from LLMlight import LLMlight

# Initialize with default settings
client = LLMlight(model='openai/gpt-oss-20b', endpoint='http://localhost:1234/v1/chat/completions')

# Run a simple query
response = client.prompt('What is the capital of France?',
                         context='The capital of France is Amsterdam.',
                         instructions='Do not argue with the information in the context. Only return the information from the context.')
print(response)
# According to the provided context, the capital of France is Amsterdam.



### 3. Using with LM Studio

```python
# Import library
from LLMlight import LLMlight

# Initialize with LM Studio endpoint
client = LLMlight(model='mistralai/mistral-small-3.2',
                  endpoint="http://localhost:1234/v1/chat/completions")

# Run queries
response = client.prompt('Explain quantum computing in simple terms')

3. Query against PDF files

# Load library
from LLMlight import LLMlight

# Initialize with default settings
client = LLMlight(model='mistralai/mistral-small-3.2',
                  context_strategy='chunk-wise',
                  retrieval_method='naive_rag',
                  embedding={'memory': 'memvid', 'context': 'bert'},
                  top_chunks=5)

# Read pdf
path = 'https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf'
pdf_text = client.read_pdf(path, return_type='text')
context = pdf_text + '\n More text can be appended in this manner'

# Make a prompt
response = client.prompt('What is an attention network?',
                         context=context,
                         instructions='Answer the question using only the information from the context. If the answer can not be found, tell that.')

print(response)

4. Global Reasoning

from LLMlight import LLMlight
client = LLMlight(model='microsoft/phi-4', context_strategy='global-reasoning')

path = 'https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf'
pdf_text = client.read_pdf(path, return_type='text')

# Make a prompt
response = client.prompt('What is an attention network?',
                         context=pdf_text,
                         instructions='Answer the question using only the information from the context. If the answer can not be found, tell that.')
print(response)

5. Creating Local Memory Database

# Load library
from LLMlight import LLMlight

# Initialize with default settings
client = LLMlight(model='mistralai/mistral-small-3.2', file_path='local_database.mp4')

url1 = 'https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf'
url2 = 'https://erdogant.github.io/publications/papers/2020%20-%20Taskesen%20et%20al%20-%20HNet%20Hypergeometric%20Networks.pdf'

# Add multiple PDF files to the database
client.memory_add(files=[url1, url2])

# Add more chunks of information
client.memory_add(text=['Small chunk that is also added to the database.',
                        'The capital of France is Amsterdam.'],
                  overwrite=True)

# Add all file types from a directory
client.memory_add(dirpath='c:/my_documents/',
                  filetypes = ['.pdf', '.txt', '.epub', '.md', '.doc', '.docx', '.rtf', '.html', '.htm'],
                  )

# Store to disk
client.memory_save()


# =============================================================================
# Load from database
# =============================================================================

# Import
from LLMlight import LLMlight
# Initialize with local database
client = LLMlight(model='mistralai/mistral-small-3.2', file_path='local_database.mp4')

# Get the top 5 chunks
client.memory_chunks(n=5)

# Search through the chunks using a query
out1 = client.memory.retriever.search('Attention Is All You Need', top_k=3)
out2 = client.memory.retriever.search('Enrichment analysis, Hypergeometric Networks', top_k=3)
out3 = client.memory.retriever.search('Capital of Amsterdam', top_k=3)

6. Load Local Memory Database

# Import library
from LLMlight import LLMlight

# Initialize with default settings
client = LLMlight(preprocessing=None, retrieval_method=None, path_to_memory="knowledge_base.mp4")

# Create queries
response = client.prompt('What do apes like?', instructions='Only return the information from the context. Answer with maximum of 3 words, and starts with "Apes like: "')
print(response)

Maintainer

• Erdogan Taskesen, github: erdogant
• Contributions are welcome.
• Yes! This library is entirely free but it runs on coffee! :) Feel free to support with a Coffee.