sura-rag 0.1.0

Verified data deletion and leak detection for RAG systems

sura-rag

Verified data deletion and runtime leak detection for RAG systems. GDPR Article 17 compliant forget pipeline with multi-strategy leak probing, runtime guardrailing, and signed compliance certificates. 100% local, zero cloud API required.

The Problem

RAG systems retrieve and present data from vector stores, but when a user exercises their GDPR Article 17 "right to be forgotten," simply deleting a document from the vector store is not enough. The LLM may have memorized fragments during retrieval, cached chunks may persist, and there is no way to verify that the data is truly gone. sura-rag closes this gap by providing a complete forget pipeline: delete → probe → guardrail → certify.

Quick Install

# Core (Ollama-based, no GPU required)
pip install sura-rag

# With CPU embeddings (sentence-transformers)
pip install sura-rag[cpu]

# With CUDA support (pre-install CUDA torch first)
pip install sura-rag[cuda]

# With framework connectors
pip install sura-rag[langchain]
pip install sura-rag[llamaindex]

# Everything
pip install sura-rag[all]

30-Second Quickstart

import sura_rag as sr

# Connect to your vector store
client = sr.SuraClient(
    vector_store=sr.adapters.ChromaDBAdapter("my_collection"),
    config=sr.SuraConfig(generator_model="llama3.2:3b"),
)

# Forget a document (GDPR Article 17)
result = client.forget(
    doc_ids=["doc_001"],
    subject="John Smith salary records",
    requestor_id="user_4821",
    regulation="GDPR_Art17",
)

print(f"Score: {result.forget_score.composite_score}")  # 0.0–1.0
print(f"Status: {result.status}")                       # "completed"
print(f"Certificate: {result.certificate_id}")          # UUID

Features

Feature	Phase 1 (v0.1)	Phase 2 (planned)
Vector store deletion	✅	✅
Fingerprint registry	✅	✅
Direct entity probes	✅	✅
Paraphrase probes	✅	✅
Contextual probes	✅	✅
Adversarial probes	✅	✅
Runtime guardrail (4 modes)	✅	✅
Audit logging (SQLite/Postgres)	✅	✅
PDF compliance certificates	✅	✅
LangChain connector	✅	✅
LlamaIndex connector	✅	✅
Parametric unlearning (LoRA)	—	✅
TOFU benchmark evaluation	—	✅
Multi-GPU training	—	✅

Architecture

SURA-RAG follows a pipeline architecture: Delete → Probe → Guardrail → Certify. Documents are deleted from the vector store, their fingerprints are stored for runtime monitoring, multi-strategy probes verify the deletion, and a compliance certificate is generated. The runtime guardrail continuously scans all RAG responses against the fingerprint registry to catch any residual leakage.

Compatibility

Component	Supported
ChromaDB	✅ ≥0.5.0
Qdrant	✅ ≥1.9.0
FAISS	✅ ≥1.8.0 (soft-delete)
LangChain	✅ ≥0.2.0
LlamaIndex	✅ ≥0.10.0
Ollama	✅ ≥0.2.0
HuggingFace	✅ ≥4.40.0
PyTorch	✅ ≥2.2.0
Pandas	✅ ≥2.0.0
Python 3.10	✅
Python 3.11	✅
Python 3.12	✅
Windows	✅
Linux	✅
macOS	✅

Local Setup

1. Install Ollama

# macOS/Linux
curl -fsSL https://ollama.com/install.sh | sh

# Windows — download from https://ollama.com

2. Pull models

ollama pull llama3.2:3b
ollama pull nomic-embed-text

3. Start Ollama

ollama serve

4. Install sura-rag

pip install sura-rag
# or for development:
git clone https://github.com/SURA-RAG/sura-rag.git
cd sura-rag
pip install -e ".[dev,cpu]"

5. Run tests

pytest tests/unit/ -v

Environment Setup

Copy .env.example to .env and fill in your values:

cp .env.example .env

The .env file is in .gitignore and will never be committed. For Phase 1 (Ollama-based), no tokens are required. See .env.example for all available settings.

Contributing

Contributions are welcome! Please:

1. Fork the repository
2. Create a feature branch: git checkout -b feature/my-feature
3. Run tests: pytest tests/unit/ -v
4. Run linting: ruff check sura_rag/
5. Submit a pull request

License

MIT License. See LICENSE for details.

Citation

If you use sura-rag in academic research, please cite:

@software{sura_rag_2024,
  title = {sura-rag: Verified Data Deletion and Leak Detection for RAG Systems},
  author = {Saxena, Aditya},
  year = {2024},
  url = {https://github.com/SURA-RAG/sura-rag},
  license = {MIT},
}