Prepare a controlled collection
Identify current documents, duplicate policies and material that must stay outside the assistant. Preserve source identifiers and version information so an answer can be traced to the document used.

Your policies and operating instructions live in documents, but employees still need help finding the relevant section. We build an internal assistant that retrieves approved material and answers with source references. The first scope uses a defined Google Drive collection and Claude, with access checks before any document text reaches the model.
Choose a specific document collection and employee audience first. The implementation must account for the source files, their permissions and the way changes reach the retrieval index.
Identify current documents, duplicate policies and material that must stay outside the assistant. Preserve source identifiers and version information so an answer can be traced to the document used.
Design identity and permission checks for the chosen collection. An index does not inherit Google Drive permissions automatically. The system must exclude denied content before generation, including when a user loses access after a document was indexed.
Connect retrieved passages to answer citations and provide a clear response when evidence is missing or conflicting. Source links support review; staff still need to check the underlying document before relying on a consequential interpretation.
Define refresh, removal and permission-change handling. Content owners need a way to retire a source and investigate unanswered questions without treating every employee conversation as a new approved policy.
Provide the Drive collection, supported file types and the people responsible for document accuracy. Identify groups that share the same access and any exceptions that require separate handling.
Confirm how employees sign in and which model account may process the material. Check data retention, logging access and available API features before indexing confidential documents.
Supply questions with known answers, conflicting versions and examples that should return no answer. Include users with different access so the evaluation covers permission boundaries as well as helpfulness.
Inspect representative files and their access paths. Agree on a supported audience and source set before implementing retrieval.
Compare cited answers against the source text. Test a denied document, revoked access, deleted source and a question whose answer is absent.
Document how new files enter the collection, how removed files leave it and who reviews content issues. Provide a way to disable a source while a problem is investigated.