RAG & Vector Security

Your AI assistant cannot intentionally answer from another organization’s knowledge base when tenant scoping is correctly enforced.

Tenant-Isolated Vector Collections

Knowledge chunks are stored in organization-specific ChromaDB collections or equivalent tenant-scoped indexes. At query time, Rumbe resolves the authenticated organization and searches only that tenant’s collection.

This boundary applies to uploaded documents, scraped help content, article metadata, and source links.

Approved Knowledge Sources

Administrators control which files, pages, and articles enter the knowledge base. Rumbe can ingest uploaded documents and crawl approved website content, then split the text into retrieval chunks and generate vector embeddings.

Customers should review content quality, permissions, freshness, duplication, and sensitive-data exposure before ingestion.

Source-Grounded Answers

When retrieval returns a relevant match, Rumbe can pass the source title and URL with the context. The response interface can display citation badges that let users inspect the supporting content.

Citations improve transparency but do not guarantee correctness. The model may still summarize, combine, or misinterpret retrieved text, so sensitive answers require human oversight.

Hidden Internal Sources

Administrators can mark an article as hidden from customer-facing sources. Rumbe may use the content as internal context while excluding its title or link from the displayed citation list.

This feature helps protect internal SOPs, escalation rules, and proprietary documents, but administrators should still avoid uploading secrets that the model is not permitted to use in responses.

Confidence Thresholds

If the highest vector match is below an approved threshold, Rumbe can decline to improvise and transfer the case to a human agent. Thresholds should be tuned using representative support queries and monitored for false confidence and unnecessary escalation.

Preventing Cross-Tenant Leakage

• —Server-derived tenant context
• —Tenant-specific collection names or namespaces
• —Authorization checks before ingestion, retrieval, deletion, and citation access
• —Tenant-aware cache keys
• —Cross-tenant negative tests
• —Separate export and backup controls
• —Logging of knowledge administration