RAGScaleGuard is intended for enterprise retrieval evaluation and diagnostics. Treat all corpora, questions, reports, and traces as potentially sensitive.

Security fixes target the current main branch until the project starts publishing tagged releases.

Please report vulnerabilities privately through GitHub Security Advisories, or by contacting the maintainers listed by the project owner. Do not open public issues for vulnerabilities involving credential leakage, prompt/report injection, unsafe data handling, or corpus data exfiltration.

Include:

• Affected file and line range
• Reproduction steps
• Impact and trust boundary crossed
• Python version and package version or commit SHA

RAGScaleGuard is a library and evaluation toolkit, not a hosted service. It does not authenticate users, store secrets, or execute model calls by default.

Security boundaries expected from integrators:

• Keep corpora and enterprise exports outside version control.
• Run evaluations in isolated environments when using proprietary corpora.
• Do not send raw documents, prompts, reports, or traces to hosted evaluators without explicit approval.
• Apply access control, audit logging, retention, and deletion policies in the host platform.
• Treat generated Markdown/JSON reports as sensitive artefacts.
• Treat diagnostic artefacts as sensitive because they can expose document IDs, source relationships, ranking metadata, and internal failure modes.

• Report output redacts common secret patterns.
• Markdown reports escape corpus-controlled fields to reduce report injection risk.
• The default package has no network dependencies.
• The model reranker is only an adapter and does not call external services unless an integrator supplies one.
• Diagnostic artefacts are structured so teams can review or redact them before exporting to observability, ticketing, or adviser systems.

• Multi-tenant isolation
• Hosted API security
• Key management
• Policy enforcement across external vector stores or model providers
• Full data loss prevention or PII classification
• Automatic security approval for third-party adviser models