Format · RAG and agent data

Keep personal data out of your RAG index

Once a document is chunked and embedded, the personal data in it can come back in any answer, to anyone who can ask. The cheapest place to remove it is before ingestion, and the safest index for development is one built from synthetic documents.

Get a 48-hour proof Keep data out of LLMs

Where Synthesized sits in the pipeline

Before anything is chunked or embedded. Not in the request path.

Source documents

Contracts, tickets, emails, notes

02 · Synthesized

Mask, or replace with synthetic

Token map stays with you

Chunk and embed

Vector index

Retrieve and answer

Traces logged for evaluation

What you can do

Engineer-ledClean source documentsMask PDF, DOCX and PPTX files before they are embedded, with consistent tokens so retrieval still links related passages.

Engineer-ledSynthetic corporaBuild development and test indexes from synthetic documents, so prototypes never touch real customers.

Engineer-ledAgent traces and memoryMask the prompts, responses and tool calls you keep for evaluation and regression tests.

What's in a typical corpus

Contracts with named partiesSupport ticketsEmailsMeeting notesHR policies with examplesPrompts and answers in logsTool calls with customer IDs

Used for

Keeping data out of LLMs in development Testing AI agents Testing PII guardrails

Questions about RAG and agent data

Why mask before embedding rather than at query time?

A filter at query time has to catch every leak, every time, in every answer. Data that never entered the index can't leak at all. Most teams want both, but only the first is under your full control.

Can I remove personal data from an index I already built?

Not reliably in place. The practical route is to mask the source documents and rebuild the index from them.

Does masking hurt answer quality?

Blunt redaction can: a model given "[REDACTED] signed with [REDACTED]" has less to go on. Consistent tokens or realistic replacements keep who-did-what intact, so retrieval and answers still work.

The platform underneath, in production

Documents run on the same platform banks and insurers already trust with their test data.

Documents and text use the same policies, the same engine and the same deployment as these results.

Banking · test data80%less time to provision test data, with compliant data generated through the API inside CI/CD.Deutsche BankCase study

Insurance · security0security waivers across 14 priority applications: 28M+ rows secured, 100% referential integrity, test data 60% faster.Global specialty insurerCase study

Published results from database and synthetic data work on the Synthesized platform.

Build your next RAG prototype on safe data

Masked documents or synthetic ones, prepared in your environment. Start with 50 of your documents in 48 hours, for a fixed fee.

Get a 48-hour proof