Connectors
Globus does not ask anyone to move their work into a new system. An employee connects their own Google account, and their files and mail become a searchable memory that is private to them. This page is the inventory: what is read, what is skipped, and what is not on the corporate portal yet.
On the corporate portal today
An employee signs in to the portal, opens the connect page, and grants access themselves. Nobody connects a mailbox on their behalf, and there is no shared corporate mailbox key behind it.
Everything in that employee’s Drive that is not in the trash and was modified in the last five years is listed and classified, and every file Globus can read is downloaded and extracted to plain text — one text file per document.
A rolling window of recent mail, with spam and trash excluded. One indexed record and one text file per message, headers kept.
Analytics ingest is live and the connect flow accepts it, so a property can be attached to a connection today. The corporate portal’s connect page currently shows only the Drive and Gmail checkboxes, so asking for Analytics is a conversation rather than a click.
The corporate portal connects Google and nothing else — the connect page is Google-only by design. The Google refresh token is encrypted at rest, and a connect flow started by one session cannot be completed by another.
Google Drive
The sync lists the Drive, decides what each file is by its type, then pulls and extracts the ones it can read — up to twenty-four files at a time. There is no interpretation step at ingest: what lands on disk is the document’s own text.
Docs export as Markdown. Sheets export as XLSX and are flattened to text, so every tab survives — a CSV export would have handed back only the first one. Slides, Drawings and Apps Script projects are exported too.
The four attachment formats people actually save and forget about, plus the plain-text families. Each is parsed on the server and written out as text.
Anyone can email a staff member a document, and that document ends up being parsed inside our server. So an Office archive that declares more than 150 MB of uncompressed content is refused before any of it is decompressed, any XML part carrying a DTD or entity declaration is refused outright, and a PDF is bounded at five thousand pages.
A file that fails to extract raises a failure and is recorded as one. It is never written to disk as though the error message were the document’s own text.
Drive · what is skipped
The practical consequence: on any real Drive, the number of files indexed is larger than the number carrying extracted text. Both are stored per file, with a reason on every skip, so the gap is inspectable rather than mysterious.
Gmail
The default is a rolling ninety days with spam and trash excluded. The window is stored per connection and read fresh on every sync, so it can be changed without a rebuild — but there is no control for it in the product today: it is a change we would make on the server. Every connection on our own deployment runs the default.
newer_than:90d -in:spam -in:trash
Each message becomes one indexed row, and every message carrying a text body also gets one text file on disk with the headers preserved — so a search result points at a real message rather than at a summary of one.
Fifty thousand messages per sync, and each body truncated at fifty thousand characters. A very long thread is read, not read whole.
The employee’s assistant can also search their live mailbox directly, so a message that arrived after the last sync is still findable — it simply is not part of the summarised memory yet.
The shared company knowledge base
Alongside each person’s private memory there is one reserved shared namespace. Every per-member search unions it in, and it is folded into every member’s context labelled plainly as shared and not private to that user. The mechanism is built and wired into both paths.
Available on the individual product, not yet on the corporate portal
These connectors are real and running on the individual product. None of them is wired on the corporate portal, and none is offered to an employee today. They are listed because a buyer deserves to know the difference between “we could” and “you can”.
| Channel | What it actually does | Corporate portal |
|---|---|---|
| Telegram | A logged-in user session mirrors messages from watched chats into a message store, which the assistant can search and which feeds the summarised memory. | Not on the corporate portal today |
| A Chrome extension reads WhatsApp Web’s own page in the browser. It records only while that tab is open, so capture stops the moment the tab is closed — and the data then goes stale quietly rather than loudly. | Not on the corporate portal today | |
| Microsoft Teams | Group chats only, deliberately skipping one-to-one and channel posts. It is ingested and counted — but there is no Teams search tool and it does not reach the summarised memory, so the assistant cannot answer from it. | Not on the corporate portal today |
| Obsidian | A vault uploaded as a zip or pasted in; the markdown is extracted and folded into that person’s memory as a single source. Unlike Drive, it is not indexed file by file, so the per-file search tools do not return individual notes. | Not on the corporate portal today |
| Freshsales CRM | Contacts, accounts and deals sync as a first-class source, not as a side note — the same search that reads a document reads a contact record. | Not on the corporate portal today |
Wiring any of these onto the corporate portal is real work, not a switch being held back behind a sales call. If one of them is the reason you would buy, say so on the call and we will tell you honestly what it would take.
Freshness
A background worker wakes every ten minutes and re-syncs any connection that has not run in the last hour. Nothing is triggered by a file being saved — the loop finds it on its next pass.
Every file in the window is re-listed and re-fetched on each run. Only the expensive summarising step is skipped, keyed on a hash of the content, so a rename or a permission change does not trigger it.
A job checks every twenty minutes, but it only rebuilds a person’s summary once their data has run roughly forty-eight hours ahead of it, and it rebuilds at most a couple of people per pass. So the condensed memory can be a day or two behind. The live search tools are not — they read current data.
Put plainly: a document saved a moment ago is very likely not indexed yet. Ask about it later in the day, or ask the assistant to search the live mailbox or Drive, which does not wait for the loop.
The inventory, against your Drive
The fastest way to know what Globus would read is to hold this page against what your team actually keeps. Tell us where the work lives, and we will tell you which parts it can read, which parts it cannot, and which of them we would have to build.