Ensure no PII is stored in vector databases or embeddings
RAG pipelines often accidentally store names, phone numbers, or identifiers in vectors. Guardrails run pre-embedding PII scanning and enforce safe ingestion.
What's at stake
- Vector databases persist your data indefinitely—personal information becomes permanently embedded
- Embeddings can leak personal data through similar document retrieval
- Deleting source documents doesn't delete the embedded representations
- GDPR right-to-erasure requests become nearly impossible to fulfill
- Enterprise customers audit your RAG architecture during security reviews
How to solve this
RAG pipelines typically chunk documents, generate embeddings, and store them in a vector database. If those documents contain personal data, that data gets embedded and persisted. The problem compounds: embeddings are opaque, making it hard to audit what personal data they contain.
The solution is to clean documents before embedding. Every chunk must be scanned for PII before it's sent to the embedding model and stored. Names, email addresses, phone numbers, and identifiers should be removed or replaced with tokens.
This must happen in your ingestion pipeline, before any data reaches the embedding model or vector store. After-the-fact cleanup is expensive and incomplete.
How Superagent prevents this
Superagent provides guardrails for AI agents—small language models purpose-trained to detect and prevent failures in real time. These models sit at the boundary of your agent and inspect inputs, outputs, and tool calls before they execute.
For RAG pipelines, Superagent's Redact model scans document chunks before they're embedded. When your ingestion pipeline processes a document, Redact inspects each chunk for personal data—names, emails, phone numbers, addresses, and custom patterns—and removes or masks them.
The embedding model and vector database only see cleaned content. Your retrieval continues to work normally, but queries never return chunks containing personal data.
Redact integrates into your existing ingestion pipeline. Point it at your chunking output, and it produces clean chunks ready for embedding. Audit logs show what was detected and removed for compliance verification.