A retrieval-augmented generation, or RAG, chatbot finds relevant material before generating an answer. For a business, the hard question is not whether a chatbot can read documents. It is whether it can answer the right people using the right version of those documents.
Define the knowledge boundary
List the approved source systems, document owners and update frequency. Mark internal, customer-specific and public material separately. Decide whether users can see the original document or only a permitted excerpt.
Explain how superseded documents are removed. Otherwise, an impressive prototype can give an answer from an old price list or a policy that no longer applies.
Specify observable behaviour
A practical brief should require answers with traceable references, a clear response when evidence is missing and a route to a person. Ask for tests that confirm one customer's material cannot appear in another customer's answer.
Prepare questions covering direct facts, ambiguous wording, conflicting documents and information absent from the collection. Record the acceptable answer and the expected source for each. Keep some questions separate from development so the final assessment is not just a replay of the demonstration.
Request a maintainable handover
Ask the developer to document ingestion, document updates, retrieval settings, access controls and failure alerts. Your team should know how to remove a source, investigate an answer and disable the service if necessary.
Compare proposals using the same data size and access requirements. An internal FAQ and a permission-sensitive client portal are different projects, even if their chat boxes look similar.
Google's RAG overview provides a reference for the retrieval pipeline. It does not remove the need for your own answer-quality and access tests. Explore AI specialists once those requirements are written.