Point your own AI at your documents

You are already using more than one AI assistant, and none of them can see your documents. GPTree can now sit behind whichever one you prefer, as a knowledge server it connects to directly.

Once connected, your assistant gains a small set of abilities: search your documents, open a passage word for word, verify that a passage is still what it was when you added it, and ask a full question that comes back with citations and a coverage report.

The part worth noticing

Your assistant can now say it does not know.

When your documents cannot support an answer, the connection returns a refusal rather than a plausible paragraph. That sounds like a limitation and it is the reason to use it. An assistant that always produces something readable is one you have to check every time; an assistant that declines when the material is not there is one you can act on when it does answer.

The same applies to partial answers. If the engine read part of a long contract rather than all of it, the response says so and names what it left out. You are told the shape of what was read, not just the conclusion.

Every claim opens

Answers come back with markers, and each marker resolves to an exact passage in a specific version of a specific document. Opening one is a single request, and the passage is checked against a fingerprint taken when the document was added. If the stored text no longer matches, you are told that too, which is a different answer from "not found" and a more useful one.

Permissions are narrower than you expect

Access is granted in pieces rather than all at once:

• Find passages. Searching does not include reading them back.
• Read passages. Opening a passage and verifying it.
• Ask questions. Generating an answer. This one spends credits, so it is granted deliberately.
• Add documents. Sending content in, and checking on it.

An assistant only ever sees the abilities you granted it. Giving something the ability to search your contracts without the ability to read them back is a real and sometimes sensible choice.

Credentials that expire

If you connect an assistant running on your own laptop, GPTree issues a short-lived credential rather than a permanent key. It lasts an hour, so a config file copied into a support thread or synced to a backup stops being useful quickly. Renewing it is the connecting app's job. Some assistants, including Claude Desktop, only send a fixed credential, so when the hour is up you swap in a new one yourself. The Knowledge API documentation shows how.

The refresh credential lasts longer and can do nothing on its own. It cannot search, read, ask, or add anything. It exists only to renew the short one, and revoking it also cancels whatever it has already issued.

For something running on a server you control, a long-lived key is still the simpler choice, and that has not changed.

Getting connected

Create a connection from Account on the Developer tab, choose only the abilities that assistant needs, and paste the credential into your assistant's settings. Connecting an assistant this way is available on Pro and Team plans. Asking questions spends credits at the same rate as anywhere else, and reading does not. The setup steps and the full list of what an assistant can call are in the Knowledge API documentation.