inAi builds knowledge assistants around an agreed collection of documents and data. Users can ask a question, inspect relevant sources and recognise when the available information does not support an answer.
Retrieval-augmented generation, or RAG, combines retrieval from a collection with a model-generated response. It is an architecture choice, not a guarantee of accuracy or confidentiality on its own.
Begin with the collection and the users
Identify the source systems, document owners, update frequency and permission model. Decide which materials are authoritative when versions disagree and what should happen when documents are removed.
A support specialist, an employee and a manager may have different access. Permissions must constrain retrieval and output on the server; hiding a document link in the interface is not an access-control strategy.
What an assistant should show
A useful interface can present an answer alongside citations and source previews. It should make uncertainty visible, ask for a missing detail when needed and avoid inventing a policy that is absent from the collection.
For example, a fictional employee can ask about an expense policy and open the exact current rule. A question requiring restricted material should not reveal that material simply because the user asks for a summary or requests a different language.
What inAi can deliver
The scope can include source connectors or ingestion, document processing, indexing, retrieval logic, role-aware filtering, answer generation, citations, feedback collection, evaluation datasets, interface integration and operational monitoring.
We also agree how a source update reaches the assistant, how deleted material leaves the active index and how the organisation investigates an unsupported answer.
Evaluate questions, not only search results
Build a test set containing answerable questions, questions with no supported answer, ambiguous requests, outdated sources, restricted sources and misleading text inside retrieved documents.
Evaluate whether the assistant retrieves relevant authorised evidence, represents it correctly, cites it accurately and abstains when appropriate. These are separate checks. A fluent answer and a plausible citation are not sufficient by themselves.
Choose the integration boundary
The assistant can be a standalone internal tool or an embedded feature in existing software. The right choice depends on where users already work and how accounts, source permissions and support should be managed.
Start with a bounded collection and user group. Expanding the corpus changes retrieval and evaluation requirements; it is not merely a larger file upload.
Questions to resolve early
Who owns the sources? Which documents are current? Are there different access levels? How often does content change? What questions matter most? Who will review incorrect answers and maintain the test set?
Those answers determine the scope more directly than selecting a model name at the start.