Knowledge base

A knowledge base lets an assistant or agent answer from your own documents, instead of only what it already knew.

Sources, documents, notes, pages, and Google Drive files

A source is a named container you create in Knowledge (e.g. "Product docs" or "Acme Corp"). A source can hold any mix of four kinds of content:

• Uploaded documents — a PDF, Word (.docx), PowerPoint (.pptx), Excel (.xlsx), plain-text (.txt), or Markdown (.md) file. A PowerPoint file is indexed slide by slide, so a search result can point you to the specific slide it came from. An Excel file is indexed row by row, with each row's column headers carried along — good for looking up a specific record, not for totals or averages across rows.
• Notes — markdown you type directly, for anything that doesn't already exist as a file (a quick clarification, a policy summary).
• Pages — point at a single URL and it's fetched and indexed; not a crawler, so it only indexes that one page, not the whole site. Use "Refresh" on a page to re-fetch it if the source has changed.
• Google Drive files — connect your workspace's Google account once (an admin does this from inside any source), then pick specific files from Google's own file picker. Only the files you explicitly pick are accessible — connecting doesn't expose your whole Drive. Google Docs/Sheets/Slides are indexed as text directly; other files (PDFs, Word docs, etc.) stored in Drive go through the same indexing as an uploaded file. Use "Refresh" to re-pull a file if it's changed in Drive.

Everything you add — whichever kind — goes through the same pipeline: it shows as processing while it's being read and indexed, then ready once it's searchable, or error if something went wrong.

Attaching it to an assistant or agent

A source isn't automatically available everywhere — you attach specific sources to a specific assistant or agent, the same way you'd attach a tool. Only the sources you've attached are ever searched; nothing else in the workspace is implicitly included.

How it's searched

Behind the scenes, each document is split into chunks and embedded for semantic search — a search for "refund policy" can match a chunk that never uses that exact phrase, as long as it's about the same thing. A search returns the few most relevant snippets, not the whole document; the assistant can expand a specific result to read more of that document if a snippet isn't enough.

Controlling when it searches

An assistant's knowledge mode controls how retrieval is triggered:

• Tool (the default) — the assistant decides for itself when to search, by calling a search tool. Good when most messages don't need the knowledge base at all.
• Always — every message automatically retrieves relevant snippets and adds them to context before the assistant replies, with no tool call and no decision for it to make.
• Hybrid — does the automatic retrieval of "Always," but also leaves the search tool available, so the assistant can go looking for something more specific if the automatic result isn't enough.

Removing content

Deleting a document or an entire source removes it from search immediately. There's no "soft delete" or recovery — re-add it if you need it back.