Products
SCRS — the AI data firewall Other Me — the workspace Compare both products
Solutions
All solutions AI Assistants Email + Projects CRM Live Chat Ticketing Flows HR-lite Control Center
By industry
Accountancy Legal Mortgage & IFA Healthcare
Company
Custom Solutions Pricing Blog About Us
Log in
SCRS login Other Me — Business Other Me — Family Get started with SCRS
Architecture · · 5 min read

Zero-Plaintext Vector Indexes in Plain English

AS

Founder & CEO, Pop Hasta Labs

From my perspective, “zero plaintext in the vector index” sounds like jargon. It isn’t. It’s a specific technical choice that makes a significant difference to how safe your client data is inside an AI platform.

What a vector index is

When an AI platform needs to search your data, it doesn’t search word-by-word. It converts your content into mathematical vectors — long lists of numbers that represent meaning — and searches by similarity. The vector index is where these representations live.

In many platforms, the vector index contains both the mathematical vectors and the readable source text. Convenient for debugging, dangerous for privacy — if the index is ever accessed, the source text is right there.

What zero plaintext means

A zero-plaintext vector index contains only the mathematical vectors. The readable source text lives in a separate encrypted system. When the AI needs to retrieve specific content, it matches vectors in the index first, then fetches the corresponding source text from the encrypted store — with authorisation checked at that step.

Specially, this means an attacker who somehow gained access to the vector index wouldn’t be able to read client documents. They’d see mathematical representations that are not trivially reversible. The readable content sits behind a separate authorisation boundary.

Why this matters for UK SMEs

For regulated practices, this property is one of the three structural guarantees that make the patent-pending SCRS architecture defensible. Specially under ICAEW confidentiality, SRA Principle 7, FCA SYSC 9 and Caldicott frameworks, the practice can demonstrate that client content is held in a way that limits exposure even if one system component is compromised.

Apart from this, for PII redaction purposes, the separation of vectors from plaintext makes it much easier to substitute placeholder tokens before content reaches the AI model — the plaintext resolution happens only on verified retrieval.

Learn more

Full architecture on the Security page. Related: What is a patent-pending AI data firewall. SCRS is self-serve — get started whenever you are ready.

AS

Abhishek Sharma

Founder & CEO of Pop Hasta Labs. Building SCRS — the patent-pending AI data firewall — and Other Me, the governed workspace built on it. Based in London.

Govern the AI your team already uses.

SCRS keeps client data out of the model — and Other Me is the whole governed workspace, built on it. No sales calls; set yourself up in minutes.