A practical approach
Retrieve answers from approved internal sources.
Prepare the knowledge sources
Identify useful documents, remove outdated or conflicting material, and agree who may access each source. Define how changes and deletions reach the assistant.
Retrieve before answering
Retrieval-augmented generation, or RAG, searches selected content and supplies relevant passages to a language model as context for an answer. This differs from relying only on information learned during model training.
Evaluate answers and references
Use representative questions to check whether the right material is retrieved, whether the answer is supported, and whether source references are useful. Include missing, conflicting, and restricted information in the evaluation.
Illustrative use case
Example scenario: internal product guidance
Sales and support staff need answers from product manuals and internal procedures, but finding the current version takes time.
A possible workflow
An authenticated assistant searches approved documents available to that employee and returns an answer with references. When sources are missing or disagree, it directs the person to the material or its owner.
What we evaluate
Employees can find the relevant guidance and inspect its source. We compare retrieval quality, answer usefulness, and search effort before deciding whether to expand the document collection.
Define the scope
Agree source ownership and access controls.
- Which questions would be useful to answer, and who asks them?
- Where are the source documents, and who maintains them?
- Which permission rules must apply when retrieving information?
- How will we handle changed, deleted, contradictory, or missing sources?
Design the boundaries
RAG can provide relevant context, but it does not guarantee a correct answer. Access restrictions must apply before restricted content reaches the model, and document content must not be treated as permission to perform actions.
From evaluation to implementation
We can propose implementation when the requirements and feasibility are clear. Otherwise, we agree on a separately priced assessment or focused evaluation first, with documented findings and a recommended scope.
Before work begins, we confirm the delivery approach, acceptance criteria, and price or budget. Hosting, model usage, and platform subscriptions are identified separately. Changes to scope or cost require your agreement. Ongoing monitoring and support are optional and scoped separately.
Discuss this AI use caseBefore we begin
Is RAG the same as training a model on our documents?
No. RAG retrieves material at query time and provides it as context. Fine-tuning changes model behavior through training. The choice depends on the task; an internal knowledge assistant does not automatically require fine-tuning.
Can it respect different employee permissions?
That must be part of the design. We review identity, source permissions, and retrieval controls, and test that users cannot retrieve information outside their access rights.