The localrag-kit maintainer team considers privacy, air-gapped security, and document isolation fundamental design constraints. Because localrag-kit indexes sensitive local codebases, internal documentation, and proprietary PDFs, we enforce strict local-first security boundaries.

Security updates are actively maintained for the following versions:

If you identify a security vulnerability (such as unauthorized outbound network telemetry, SQLite/FTS5 injection vectors, local file traversal during ingestion, or model prompt injection risks):

1. Do NOT report via public GitHub issues. Public issues expose user environments to risk before a defensive patch can be distributed.
2. Use GitHub Private Vulnerability Reporting:
   • Navigate to the Security tab on GitHub.
   • Click "Report a vulnerability" to submit a private security advisory.
3. Alternative Direct Email:
   • Email our security maintainers directly at: security@ritualdev.com
   • Please include:
     • Clear description of the vulnerability and affected subsystem (localrag.storage, localrag.indexer, localrag.server).
     • Reproduction steps or proof-of-concept (PoC).
     • Operating system, Python version, and execution environment.

• Initial Triage: Within 24-48 hours.
• Risk Assessment: Within 72 hours.
• Coordinated Disclosure: Security patches will be published to PyPI immediately upon fix validation.

• 100% Air-Gapped & Offline: LocalRAG-Kit operates entirely on local compute. Core indexing and hybrid search make zero background network requests, zero telemetry calls, and zero external vector DB syncs.
• Parameterization Against SQL Injection: All queries to SQLite and the FTS5 full-text search virtual table strictly use parameterized queries (?). User query strings are never concatenated directly into raw SQL.
• Safe Path Traversal Protection: Ingestion paths are strictly resolved against verified local directories, preventing symlink traversal attacks outside designated repository boundaries.
• Model Sandbox: When running local models via Ollama or local GGUF runtimes, prompts and context chunks are passed through structured templates without shell evaluation.