How to prepare an AI assistant’s information
An AI assistant needs approved information, clear conditions and access limits. Uploading documents is only one step. This guide shows how to turn scattered information into a knowledge base that can answer, present the source and route questions that need a person.
By the end, you can:
- Organise sources and their owners
- Define permissions and handover
- Test answers before making the assistant available
Start with the questions the assistant should resolve
Choose a specific task: explain services, help someone use a product or look up information about an authenticated order. List frequently asked questions, permitted decisions and matters requiring handover. Answering about a service, accessing an account and changing an order require different capabilities and authorisations.
In a RAG architecture, the system finds relevant information and provides that context to the model to prepare the answer. This is not the same as training the model on all the documents. Microsoft’s documentation describes this process. The method helps consult your own sources, but still requires tests of retrieval, answers and access.
Assign an owner and validity conditions to every source
Create an inventory before importing. Record the subject, owner, approval status, authorised audience, version and validity conditions. A document without an owner can remain technically accessible after it no longer represents operations.
Define precedence by subject. For commercial terms, the source approved for that purpose takes precedence over an old presentation. The file’s modification date alone does not resolve a conflict. If two approved sources disagree, suspend the affected answer and ask the responsible person to decide.
| Source | Editorial decision | Access |
|---|---|---|
| Approved catalogue | Index with version and owner | Public, if the catalogue is public |
| Ended campaign | Remove from current commercial information | Separate archive, if necessary |
| Internal procedure | Confirm approval and scope | Authorised working group |
| Order data | Look up in the responsible system | Authenticated and authorised customer |
Write complete units of knowledge
Each unit should keep the answer together with the conditions and exceptions that make it correct. Preserve headings, lists and relationships between table cells. If a PDF was scanned, check that values and symbols have been read correctly after extraction. An extraction error can turn a correct source into a wrong answer.
When dividing documents into chunks, keep enough context to interpret each passage. Microsoft’s chunking documentation presents strategies based on structure and meaning. The right size depends on the documents and tests; there is no word count that solves every case.
id: atendimento-prazo-orcamento
estado: aprovado
versao: 3
publico: publico
responsavel: operacoes
pergunta: When will I receive the quote?
resposta: The timescale is confirmed after the request has been assessed.
condicao: The equipment and the work must be known.
excecao: Incomplete requests need clarification.
fonte: procedimento-atendimento-v3
rever_em: 2026-12-17Apply permissions before retrieving information
The system must establish identity and access through the application. Writing “do not show private data” in the assistant’s instructions is not enough. Documents eligible for retrieval must respect the permissions of the person asking the question, including chunks derived from each document.
Microsoft’s document-level access-control documentation shows this distinction between content and authorisation. In a simple project, separating the public knowledge base from the restricted area may make operations easier. In any design, test that a user cannot retrieve documents from another account, even when they know the title or identifier. Permission changes must also reach the index.
Prepare evidence-based answers with a useful next step
Ask for answers supported by approved passages, with a reference the user can consult when authorised. A correct link does not, by itself, prove the conclusion is correct: the source must support the claim. When conditions are missing, the assistant asks for clarification; when evidence is missing, it identifies the gap and routes the request onwards.
OWASP describes prompt injection: instructions received in a question or hidden in a document may try to divert the model’s behaviour. Treat retrieved documents as content, not as authorisation to carry out actions. The application must enforce permissions, limit operations and require human validation when the decision warrants it. A security instruction on its own does not eliminate this risk.
- Authorise
Determine the information the person can access.
- Retrieve
Find approved sources valid for the subject.
- Answer
Explain the conclusion and identify the evidence.
- Hand over
Ask for context or pass to a person when necessary.
Handover should preserve the necessary context without exposing information that should not be disclosed.
Test what it should answer and what it should recognise as missing
Create a set of questions with the expected source, mandatory conditions and acceptable behaviour. Add different wordings and repeat critical cases: a single correct answer does not demonstrate consistency. Separate retrieval failures, outdated content, interpretation and authorisation to fix the root of the problem.
| Test case | Expected result |
|---|---|
| Question covered by the knowledge base | Answer consistent with the passage and conditions. |
| Ambiguous question | Clarification before assuming a condition. |
| Missing information | Gap acknowledged and useful handover. |
| Two contradictory sources | Conflict handled according to approved precedence. |
| Document from another account | Content excluded from retrieval and the answer. |
| Unauthorised instruction in a document | No change to authorisation and no unauthorised action. |
Maintain the knowledge with the same care as the service
An update should go through review, source publication, index update and rerunning the affected tests. Confirm that the old version is no longer retrieved when it no longer applies. Record the change and prepare to restore the previous approved version if the update introduces errors.
Also review questions handed over to a person: they may reveal a missing explanation, a confusing process or a new need. Add knowledge after validating it with the responsible person. The aim is to improve coverage while keeping answers justifiable.
Worksheet for your project
Assistant knowledge worksheet Assistant’s task and audience: Questions it should resolve: Subjects and actions requiring handover: Approved source for each subject and its owner: Version, validity and next review: Conditions and exceptions accompanying each answer: Groups or accounts authorised to consult the source: Rule for contradictory sources: Expected answer when information is missing: Test cases and acceptance criteria: How to update the index and remove old versions: Handover channel and information to pass on: Person responsible for operations and approval date:
Sources and criteria
Content prepared with AI assistance and research using technical sources, with checks of the relevant examples and calculations. The references support the guide’s decisions; examples identified as fictitious let you try out the method.
