RAG Ingestion Sanitization Sandbox
Paste or edit a sample document below to see why pattern-based classification has to run before a chunk is embedded, not after. The sandbox flags sensitive spans, redacts them inline, and shows what an embedding pipeline would actually ingest.
The classification step below runs real pattern matching in your browser, the same category of regex-based detection covered in the vector boundary guide, applied to whatever text you provide. The "vector" representation shown after sanitization is an illustrative placeholder, a deterministic visual stand-in for what an embedding looks like, not output from an actual embedding model. No text you enter is sent anywhere; everything runs locally in this page.
The chunk size, overlap, and embedding model selections shape how text is split for this demonstration; they do not change which embedding model actually classifies anything, since no real embedding is computed here. Detection logic below covers a small illustrative set of patterns (SSNs, emails, mock credit card numbers, a flagged internal codename). Production classifiers use far broader pattern libraries plus ML-based context classification, as covered in the vector boundary guide.