GuidesAI and agents

An internal AI assistant connected to Drive and Notion

By the Banzai Labs teamUpdated on 3 min read

DriveNotionWikiTicketsPermissions ofthe userSource: HR procedure.pdfIt only reads what the person is allowed to see
The assistant queries Drive, Notion and the wiki with the person’s permissions, and cites its source.

Two architectures

Search on demand, or index
Live
  • The assistant calls the Drive and Notion APIs
  • With the person's account
  • Simple, always up to date
Index
  • Documents chunked and embedded
  • Permissions copied onto each chunk
  • Fast on large volumes

The building blocks

Block Role
Connectors Read Drive, Notion, the wiki, through their API or an MCP server
Authentication Each user signs in with their own account
Search By keywords, by meaning, or both
Model Answers from the excerpts found, citing the document

Rules for a good answer

  1. 1A cited sourceTitle and link to the document.
  2. 2Nothing made up“I couldn't find it” rather than a guess.
  3. 3The latest documentThe modification date settles two versions.

Before launch

  • Clean up: old, contradictory documents produce bad answers.
  • Test with twenty real questions from the team.
  • Measure: questions asked, useful answers, “I couldn’t find it”.

The method: RAG on your documents. Connectors: the MCP server.

Frequently asked questions

Do I have to copy everything into a vector database?
Not necessarily. For a small volume, querying the sources live with the user's permissions is enough. An index becomes useful when there are many documents.
How do I respect access rights?
Either the assistant queries Drive and Notion with the person's own account, or the index keeps the list of allowed people for each document and filters before answering.
Is there a ready-made option?
Yes: Claude, for example, offers connectors to Google Drive, Notion and other tools in its team plans. Custom work makes sense for internal sources or special rules.

Read next

All guides