RAG Knowledge Assistant: Setting Safe Data Boundaries

Retrieval-augmented generation (RAG) lets an AI assistant answer questions using your own documents: policies, procedures, product information, contracts, or support history. It is one of the most practical uses of AI in business. It is also one of the easiest ways to expose information to the wrong people, because an assistant that can search everything will happily summarise anything. Data boundaries, decided before indexing starts, are what make a RAG assistant useful and safe. This guide explains how to set them.

What It Is

A RAG system has three main stages. Ingestion collects documents from approved sources, splits them into chunks, and stores them in a search index, often a vector database. Retrieval finds the chunks most relevant to a user's question. Generation passes those chunks to a language model, which writes an answer, ideally citing its sources. Data boundaries apply at every stage. At ingestion, you decide which sources and document types are allowed, and which are excluded. At retrieval, the system must respect the same permissions users have in the original systems; if a user cannot open a document in the file share, the assistant should not retrieve it for them. At generation, you decide what the model may do with retrieved content and what it should decline. Around all of this sit retention rules for indexes, prompts, and logs, and the policies of any external model provider about how submitted data is handled.

Who It Is For

This guide is for:

Organisations planning an internal assistant for staff knowledge, HR, IT, or operations
Teams building a customer-facing assistant on product or support content
Data protection and security leads reviewing a proposed RAG project
Managers deciding which content should and should not be searchable by AI

What It Helps Clarify

Setting boundaries helps clarify:

Which knowledge sources are approved, owned, and current enough to be trusted
Whether retrieval honours the permissions of each user
Where personal, confidential, or regulated data might enter the index
How long indexed data, prompts, and answers are kept, and where

What's Included

List candidate sources and give each an owner who approves inclusion
Exclude by default: HR case files, legal matters, payroll, health data, credentials, and board papers, unless there is a specific, approved need
Remove or archive outdated documents so the assistant does not cite superseded policies
Carry document permissions into the index and filter retrieval by the current user's access
Test with users in different roles to confirm they only see what they should
Detect and handle personal data and secrets during ingestion
Require answers to cite sources and allow users to open the original document
Instruct the assistant to say when it does not know, rather than guess
Define retention for the index, prompts, and logs, and how deletions in source systems propagate
Review the model provider's data handling and retention terms
Collect feedback on wrong or unhelpful answers and assign someone to act on it

What's Not Included

A claim that the assistant's answers will always be accurate
Legal advice on data protection obligations
Indexing of sources that have not been approved by their owners
Certification or regulatory approval

When to Choose It

Set boundaries at the very beginning, during source selection, not after the first demo. Adding permission filtering later is harder than designing it in, and a single mistaken answer that reveals confidential information can damage trust in the whole project. Revisit boundaries when new sources, user groups, or model providers are added.

Our Approach

Start small and well-governed. Choose a limited set of approved, current sources with a clear owner and audience, such as an IT help library or a product knowledge base. Build permission-aware retrieval from the start, even if the first audience is broad. Test with realistic questions from different roles and include questions the assistant should refuse. Our Internal Knowledge Assistant service prepares an assistant around approved knowledge sources, access boundaries, retrieval reliability, and controlled staff use. AI Governance, Policy & Data Safety can define the usage rules and data boundaries around it, and an LLM & AI Agent Security Review can examine retrieval risks and data exposure in an existing assistant. Measure success by trusted, cited answers to common questions, not by how many documents are indexed.

Internal Knowledge Assistant

View the Internal Knowledge Assistant service page