ragkit-local

A modular, pip-installable reliability layer for Retrieval-Augmented Generation (RAG) pipelines.

Why ragkit?

• 100% Fully Local: Requires zero external API calls, zero hosted LLM inference, and runs completely offline.
• Privacy-First: No data leaves your machine; uses local NLP rules and lightweight, open-source sentence-transformers NLI models.
• Framework-Agnostic: Plugs directly into LangChain, LlamaIndex, or any custom pythonic RAG pipeline.
• Import-Only-What-You-Need: Extremely fast import times with lazy loading of all underlying model dependencies on first invocation.

Installation

# Install the library in editable/dev mode or directly
pip install ragkit-local

# Download the required local English spaCy pipeline
python -m spacy download en_core_web_sm

Quickstart Examples

1. Query Decomposition

Split compound queries into atomic questions before retrieval.

from ragkit import QueryDecomposer

decomposer = QueryDecomposer()
sub_questions = decomposer.decompose("What is climate change and how does it affect oceans?")
print(sub_questions)
# Output: ['What is climate change?', 'How does it affect oceans?']

2. Confidence Scoring (Hallucination Detection)

Verify if the generated answer is supported by the retrieved document chunks.

from ragkit import ConfidenceScorer

scorer = ConfidenceScorer()  # Uses tiny 100MB model by default
chunks = ["Photosynthesis uses sunlight to convert water and CO2 into oxygen and glucose."]
answer = "Plants convert carbon dioxide and water into glucose using sunlight. They also produce helium."

result = scorer.score(answer, chunks)
print(f"Score: {result['score']}/100 | Verdict: {result['verdict']}")
# Output: Score: 50/100 | Verdict: partially_grounded
print("Reasoning:", result["reasoning"])

# For better accuracy, use the larger model:
# scorer = ConfidenceScorer(model_name="cross-encoder/nli-deberta-v3-base")