The index is the leak.
Most "on-prem-friendly" AI products keep your file on your server — and ship an index of it to their cloud. This page explains why that index is, for privacy purposes, a second copy of the document.
KEYWORDS
Topics, names, projects, clients, and legal matters — right there in plaintext.
CHUNKS
Verbatim passages copied straight out of the file. Not derived — the content itself.
EMBEDDINGS
Encode the meaning — and are reversible enough to reconstruct the source text.
METADATA
Filenames, authors, dates, case IDs, and paths — a map of what you hold.
RELATIONSHIPS
Which people, entities, and documents connect — disclosure without a single sentence.
A cloud index of local content is exfiltration the moment it holds enough derived information to disclose, infer, search, or reconstruct protected content outside your walls. DPLYD keeps the index on the same side of the boundary as the data.
The embeddings are reversible
This isn't theoretical. Research on embedding inversion — e.g. "Text Embeddings Reveal (Almost) As Much As Text" (Morris et al., EMNLP 2023) — reconstructs source text from embedding vectors with high fidelity. If your embeddings sit in a vendor's cloud, treat the text as being there too.
Five questions to ask any "compliant" AI vendor
▸ Where is the index stored?
▸ Where are embeddings computed?
▸ Where do chunks live at rest?
▸ What metadata leaves our network?
▸ Can you show us the data-flow diagram?